LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV...
Transcript of LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV...
![Page 1: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/1.jpg)
LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERY
Sofia Triantafillou and IoannisTsamardinosComputer Science Department, University of CreteJune 2016
![Page 2: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/2.jpg)
1. Integrative causal discoveryi. Motivation.
ii. Causal models.
iii. m-separation.
iv. Reverse engineering causal models (single data set).
v. Problem formulation: Reverse engineering causal models from multiple heterogeneous data sets.
vi. Modeling interventions/selection.
2. Logic-based causal discoveryi. Converting path constraints to logic formulae.
ii. Problem complexity.
iii. Conflict resolution.
iv. Existing algorithms.
v. Reasoning with logic based causal discovery.
vi. Non-trivial inferences-validation.
OUTLINE
2
![Page 3: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/3.jpg)
3
HETEROGENEOUS DATA SETS MEASURING THE SAME SYSTEM UNDER STUDY
Variables
Study
Thrombosis Contraceptives Protein C Breast Cancer Protein Y Protein Z
1
observational data
Yes No 10.5 Yes - -
No Yes 5.3 No - -
- -
No Yes 0.01 No - -
2
observational data
- - - Yes 0.03 9.3
- - -
- - - No 3.4 22.2
3
experimental data
No No 0 (Control) No 3.4 -
Yes No 0 (Control) Yes 2.2 -
- -
Yes Yes 5.0 (Treat.) Yes 7.1 -
No Yes 5.0 (Treat.) No 8.9 -
4
experimental data
No No (Ctrl) - - - -
No No (Ctrl) - - - -
- - - -
Yes Yes(Treat) - - - -
![Page 4: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/4.jpg)
“… In the randomized control trial, women taking contraceptives had 30% more chances of being diagnosed with thrombosis … ”
4
ISOLATED ANALYSISAnalyze data Publish results
“…Drugs reducing protein C reduced the probability of Breast Cancer and lowered the levels of Protein E…”
“…Protein E is a risk factor for Breast Cancer…”
“…The use of contraceptives is correlated with Thrombosis, negatively correlated with Breast Cancer and levels of Protein E …”
![Page 5: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/5.jpg)
5
INTEGRATIVE CAUSAL ANALYSIS
Data can not be pooled together:
Missing variables cannot be treated as missing values.
They come from different experimental/sampling conditions (different distributions).
![Page 6: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/6.jpg)
6
INTEGRATIVE CAUSAL ANALYSIS
Data come from the
same causal mechanism.
Breast Cancer
Protein C
ContraceptivesThrombosis
Protein Z
Protein E
Data can not be pooled together:
Missing variables cannot be treated as missing values.
They come from different experimental/sampling conditions (different distributions).
![Page 7: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/7.jpg)
7
INTEGRATIVE CAUSAL ANALYSIS
Identify the causal graphs
that simultaneously
fit all data.
Breast Cancer
Protein C
ContraceptivesThrombosis
Protein Z
Protein E
![Page 8: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/8.jpg)
1. Integrative causal discoveryi. Motivation.
ii. Causal models.
iii. m-separation.
iv. Reverse engineering causal models (single data set).
v. Problem formulation: Reverse engineering causal models from multiple heterogeneous data sets.
vi. Modeling interventions/selection.
2. Logic-based causal discoveryi. Converting path constraints to logic formulae.
ii. Problem complexity.
iii. Conflict resolution.
iv. Existing algorithms.
v. Reasoning with logic based causal discovery.
vi. Non-trivial inferences-validation.
OUTLINE
8
![Page 9: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/9.jpg)
9
SEMI MARKOV CAUSAL GRAPHS
𝑋 directly causes 𝑍𝑋 and 𝑌 share a latent common cause
Semi Markov Causal Graph G
• Directed edges represent direct causal relationships.
• Bi-directed edges represent confounding (latent confounders).
• Both types of edges allowed for a single pair of variables.
• No directed cycles (no causal feedback).
X
Y Z
![Page 10: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/10.jpg)
10
SEMI MARKOV CAUSAL GRAPHS
𝑋 directly causes 𝑍𝑋 and 𝑌 share a latent common cause
Joint Probability Distribution PSemi Markov Causal Graph G
Z
X YYes No
Yes Yes 0,01 0,04
Yes No 0,01 0,04
No Yes 0,000045 0,044955
No No 0,000855 0,854145
• Joint probability distribution entails conditional (in) dependencies.
• 𝐼𝑛𝑑 𝑋, 𝑌 𝒁 : 𝑃 𝑋 𝑌, 𝒁 = 𝑃(𝑋|𝒁)
•𝐷𝑒𝑝 𝑋, 𝑌 𝒁 : 𝑃 𝑋 𝑌, 𝒁 ≠ 𝑃(𝑋|𝒁)
X
Y Z
• Directed edges represent direct causal relationships.
• Bi-directed edges represent confounding (latent confounders).
• Both types of edges allowed for a single pair of variables.
• No directed cycles (no causal feedback).
![Page 11: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/11.jpg)
11
EXAMPLE OF CONDITIONAL (IN) DEPENDENCE
Data measuring: Smoking, Yellow Teeth,Nicotine Levels.
P(Nicotine Levels | Yellow Teeth) ≠P(Nicotine Levels)
P(Nicotine Levels | Smoking, Yellow Teeth) = P(Nicotine Levels|Smoking)
![Page 12: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/12.jpg)
12
SEMI MARKOV CAUSAL GRAPHS
𝑋 directly causes 𝑍𝑋 and 𝑌 share a latent common cause
Causal assumptions
X
Y Z
• Joint probability distribution entails conditional (in) dependencies.
• 𝐼𝑛𝑑 𝑋, 𝑌 𝒁 : 𝑃 𝑋 𝑌, 𝒁 = 𝑃(𝑋|𝒁)
•𝐷𝑒𝑝 𝑋, 𝑌 𝒁 : 𝑃 𝑋 𝑌, 𝒁 ≠ 𝑃(𝑋|𝒁)
Joint Probability Distribution PSemi Markov Causal Graph G
• Directed edges represent direct causal relationships.
• Bi-directed edges represent confounding (latent confounders).
• Both types of edges allowed for a single pair of variables.
• No directed cycles (no causal feedback).
Z
X YYes No
Yes Yes 0,01 0,04
Yes No 0,01 0,04
No Yes 0,000045 0,044955
No No 0,000855 0,854145
![Page 13: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/13.jpg)
13
CAUSAL ASSUMPTIONS
Causal Markov Assumption:Every variable is independent of its non-effects given its direct causes.
X
Y Z
![Page 14: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/14.jpg)
14
CAUSAL ASSUMPTIONS
Causal Markov Assumption:Every variable is independent of its non-effects given its direct causes.
𝐼𝑛𝑑(𝑌, 𝑍 |𝑋)
X
Y Z
![Page 15: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/15.jpg)
15
CAUSAL ASSUMPTIONS
𝐼𝑛𝑑(𝑌, 𝑍 |𝑋)
Causal Faithfulness Assumption:Independences stem only from the causal structure, not the parameterization of the distribution.
X
Y Z
Causal Markov Assumption:Every variable is independent of its non-effects given its direct causes.
![Page 16: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/16.jpg)
16
CAUSAL ASSUMPTIONS
𝐼𝑛𝑑(𝑌, 𝑍 |𝑋)
𝐷e𝑝(𝑌, 𝑍 | ∅)𝐷𝑒𝑝 𝑋, 𝑍 ∅)𝐷𝑒𝑝(𝑋, 𝑍 | 𝑌)𝐷𝑒𝑝(𝑌, 𝑋 | ∅)𝐷𝑒𝑝(𝑌, 𝑋 | 𝑍)
Causal Faithfulness Assumption:Independences stem only from the causal structure, not the parameterization of the distribution.
X
Y Z
Causal Markov Assumption:Every variable is independent of its non-effects given its direct causes.
![Page 17: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/17.jpg)
17
CAUSAL ASSUMPTIONS
𝐼𝑛𝑑(𝑌, 𝑍 |𝑋)
𝐷e𝑝(𝑌, 𝑍 | ∅)𝐷𝑒𝑝 𝑋, 𝑍 ∅)𝐷𝑒𝑝(𝑋, 𝑍 | 𝑌)𝐷𝑒𝑝(𝑌, 𝑋 | ∅)𝐷𝑒𝑝(𝑌, 𝑋 | 𝑍)
Causal Faithfulness Assumption:Independences stem only from the causal structure, not the parameterization of the distribution.
All independencies in the joint probability distribution can be identified in G using the graphical criterion of m-separation.
X
Y Z
Causal Markov Assumption:Every variable is independent of its non-effects given its direct causes.
![Page 18: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/18.jpg)
1. Integrative causal discoveryi. Motivation.
ii. Causal models.
iii. m-separation.
iv. Reverse engineering causal models (single data set).
v. Problem formulation: Reverse engineering causal models from multiple heterogeneous data sets.
vi. Modeling interventions/selection.
2. Logic-based causal discoveryi. Converting path constraints to logic formulae.
ii. Problem complexity.
iii. Conflict resolution.
iv. Existing algorithms.
v. Reasoning with logic based causal discovery.
vi. Non-trivial inferences-validation.
OUTLINE
18
![Page 19: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/19.jpg)
19
𝒎-SEPARATION
A path 𝑋1, … , 𝑋𝑛 between 𝑋1 and 𝑋𝑛 is 𝒎-connecting given 𝑽 if for every triple 𝑋𝑖−1, 𝑋𝑖 , 𝑋𝑖+1 on the path:
• If Χi−1 ∗→ 𝑋𝑖 ←∗ 𝑋𝑖+1 (colliding triplet), 𝑋𝑖 or one of its descendants ∈ 𝑽
• Otherwise, 𝑋𝑖 ∉ 𝑽
𝒎-connecting path => information flow => dependence
No 𝒎-connecting path => no information flow => independence (𝑚-separation)
Colliders Χi−1 ∗→ 𝑋𝑖 ←∗ 𝑋𝑖+1 are special and create an asymmetry that will allow us to orient causal direction.
![Page 20: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/20.jpg)
20
𝒎-SEPARATION
⇔ 𝐷𝑒𝑝(𝑌, 𝑍|∅)
X
Y Z
XY Z
is 𝑚-connecting given ∅
![Page 21: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/21.jpg)
21
𝒎-SEPARATION
X
Y Z
⇔ 𝐷𝑒𝑝(𝑌, 𝑍|∅)
X
Y Z
XY Z
is 𝑚-connecting given ∅
⇔ 𝐼𝑛𝑑(𝑌, 𝑍|𝑋)
XY Z
is NOT 𝑚-connecting given 𝑋
![Page 22: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/22.jpg)
22
𝒎-SEPARATION
X
Y Z
⇔ 𝐼𝑛𝑑(𝑌, 𝑍|∅)
X
Y Z
XY Z
is NOT 𝑚-connecting given ∅
⇔ 𝐷𝑒𝑝(𝑌, 𝑍|𝑋)
XY Z
is 𝑚-connecting given 𝑋
![Page 23: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/23.jpg)
23
CAUSAL MODELLING
Conditional (in)dependencies
(expected) in the joint probability distribution
Paths (m-separations/connections)
in the causal graph
G:
A
C
DB
E
Data set 𝐷measuring a
set of variables
A B C D E𝑨, 𝑩|𝑬, 𝑪 Ind
𝑨, 𝑩|∅ Dep
… …
𝑬, 𝑪|𝑨, 𝑩, 𝑪 Dep
![Page 24: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/24.jpg)
1. Integrative causal discoveryi. Motivation.
ii. Causal models.
iii. m-separation.
iv. Reverse engineering causal models (single data set).
v. Problem formulation: Reverse engineering causal models from multiple heterogeneous data sets.
vi. Modeling interventions/selection.
2. Logic-based causal discoveryi. Converting path constraints to logic formulae.
ii. Problem complexity.
iii. Conflict resolution.
iv. Existing algorithms.
v. Reasoning with logic based causal discovery.
vi. Non-trivial inferences-validation.
OUTLINE
24
![Page 25: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/25.jpg)
25
REVERSE ENGINEERING
causal graph?Data set 𝐷measuring a
set of variables
A B C D E 𝐺:
A
C
DB
E
?
???
?
??? ?
??
![Page 26: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/26.jpg)
26
REVERSE ENGINEERING
Find the (in)dependencies using statistical tests.
causal graph?Data set 𝐷measuring a
set of variables
A B C D E 𝐺:
A
C
DB
E
?
???
?
??? ?
??A B C D E𝑨, 𝑩|𝑬, 𝑪 Ind
𝑨, 𝑩|∅ Dep
… …
𝑬, 𝑪|𝑨, 𝑩, 𝑪 Dep
![Page 27: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/27.jpg)
27
REVERSE ENGINEERING
Find a graph that satisfies the implied m-
connections/separations.
Data set 𝐷measuring a
set of variables
A B C D E 𝐺:
A
C
DB
E𝑨, 𝑩|𝑬, 𝑪 Ind
𝑨, 𝑩|∅ Dep
… …
𝑬, 𝑪|𝑨, 𝑩, 𝑪 Dep
Find the (in)dependencies using statistical tests.
![Page 28: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/28.jpg)
28
MARKOV EQUIVALENCE
A
B
D
C E
A
B
D
C E
A
B
D
C E
…
• More than one graphs entail the same set of conditional independencies.• The graphs have some common features (edges/orientations).• For some types of causal graphs, Markov equivalence classes share the same
skeleton.• not semi-Markov causal graphs
![Page 29: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/29.jpg)
A
B
D
C E
29
CAUSAL DISCOVERY
A
B
D
C E
A B C D E
DataCausal graph(s)(In)dependencies paths
…
Sound and complete algorithms take as input a data set and output a summary of all the graphs that satisfy all identified conditional independencies.
![Page 30: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/30.jpg)
1. Integrative causal discoveryi. Motivation.
ii. Causal models.
iii. m-separation.
iv. Reverse engineering causal models (single data set).
v. Problem formulation: Reverse engineering causal models from multiple heterogeneous data sets.
vi. Modeling interventions/selection.
2. Logic-based causal discoveryi. Converting path constraints to logic formulae.
ii. Problem complexity.
iii. Conflict resolution.
iv. Existing algorithms.
v. Reasoning with logic based causal discovery.
vi. Non-trivial inferences-validation.
OUTLINE
30
![Page 31: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/31.jpg)
A
B
D
C E
31
INTEGRATIVE CAUSAL DISCOVERY
A
B
D
C E
A B C D E
Data sets measuring overlapping variable sets
under intervention/selection.
Causal graph(s)that
simultaneously fit all data.
…A B DC E
CA B D
![Page 32: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/32.jpg)
A
B
D
C E
32
INTEGRATIVE CAUSAL DISCOVERY
A
B
D
C E
A B C D E
Data sets measuring overlapping variable sets
under intervention/selection.
Causal graph(s)that
simultaneously fit all data.
…A B DC E
CA B D
• Every data set imposes some constraints.• Observational data impose m-separation/m-connection constraints on the candidate graph.• Experimental data?• Data sampled under selection?
![Page 33: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/33.jpg)
1. Integrative causal discoveryi. Motivation.
ii. Causal models.
iii. m-separation.
iv. Reverse engineering causal models (single data set).
v. Problem formulation: Reverse engineering causal models from multiple heterogeneous data sets.
vi. Modeling interventions/selection.
2. Logic-based causal discoveryi. Converting path constraints to logic formulae.
ii. Problem complexity.
iii. Conflict resolution.
iv. Existing algorithms.
v. Reasoning with logic based causal discovery.
vi. Non-trivial inferences-validation.
OUTLINE
33
![Page 34: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/34.jpg)
34
INTERVENTIONS (MANIPULATIONS)
Values of the manipulated variable are set solely by the intervention proceduree.g. value of a knockout gene in a knockout mice is set to zero; randomized variable in a randomized control trial.
No Junk food
Heart disease
Junk food
![Page 35: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/35.jpg)
Manipulated SMCM 𝑆𝐵
35
INTERVENTIONS
Graph (SMCG) G
A
C
DB
E• If you know the causal model, you can
model interventions.
• Values of B are set solely by the intervention procedure: If you know direct causal relations, remove all edges into the manipulated variable.
• This procedure is called graph surgery.• The resulting graph is called the
manipulated graph (symb. 𝐺𝐵)
Manipulated SMCG 𝐺𝐵
(after graph surgery)
A
C
DB
E
![Page 36: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/36.jpg)
Dataset 𝐷𝑖 measuring a subset of variables, some of
which are manipulated
Conditional independencies in 𝐷𝑖
Path constraints on the causal graph after
manipulation
𝐺𝐵:
A
C
DB
E
36
CAUSAL DISCOVERY WITH INTERVENTIONS
∄ m-connecting path from A to D given ∅ in 𝐺𝐵
∄ m-connecting path from A to D given 𝐵 in 𝐺𝐵
⋮∄ m-connecting path from A to D given B, 𝐶 in 𝐺𝐵
⋮∃ m-connecting path from B to C given ∅ in 𝐺𝐵
A DC EB 𝑨, 𝑩|𝑬, 𝑪 Ind
𝑨, 𝑩|∅ Dep
… …
𝑬, 𝑪|𝑨, 𝑩, 𝑪 Dep
![Page 37: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/37.jpg)
37
SELECTION BIAS
• Samples are selected based on the value of one of your variables.
• e.g. you perform your study in a specific region/on the internet; case-control study for a rare disease.
population
Sample (internet users)
evolution creation
![Page 38: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/38.jpg)
Manipulated SMCM 𝑆𝐵
38
SELECTION BIAS IN CAUSAL MODELS
Selected Graph (SMCG) 𝐺𝐷
• If you know the causal model, you can model selection bias.
• Samples are selected based on the value of D; The value of D directly affects the probability of being selected.
• S is a child of D, S=1 for all your samples.
• Selected graph, symb. 𝐺𝐷
Manipulated SMCM 𝑆𝐵Graph (SMCG) G
A
C
DB
E
A
C
DB
E
S=1Selection variable
![Page 39: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/39.jpg)
A
C
DB
E
S=1
Dataset 𝐷𝑖 measuring a subset of variables, some of
which are selected upon
Conditional independencies in 𝐷𝑖
Path constraints on the underlying causal graph
after selection
39
CAUSAL DISCOVERY WITH SELECTION BIAS
∄ m-connecting path from A to D given ∅ in 𝐺𝐷
∄ m-connecting path from A to D given 𝐵 in 𝐺𝐷
⋮∄ m-connecting path from A to D given B, 𝐶 in 𝐺𝐷
⋮∃ m-connecting path from B to C given ∅ in 𝐺𝐷
A C EB 𝑨, 𝑩|𝑬, 𝑪,S=1 Ind
𝑨, 𝑩|S=1 Dep
… …
𝑬, 𝑪|𝑨, 𝑩,D,S=1 Dep
D
𝐺𝐷:
![Page 40: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/40.jpg)
A
B
D
C E
40
INTEGRATIVE CAUSAL DISCOVERY
A
B
D
C E
A B C D E
Data Causal graph(s)(In)dependencies paths
…A B DC E
CA B D
• Every data set imposes some constraints.• Observational data impose path constraints on the candidate graph.• Experimental data impose path constraints on the candidate graph after manipulation.• Data sampled under selection impose path constraints on the candidate graph after selection.• Easily handles overlapping variable sets
• Each study imposes constraints on the observed variables.
![Page 41: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/41.jpg)
1. Integrative causal discoveryi. Motivation.
ii. Causal models.
iii. m-separation.
iv. Reverse engineering causal models (single data set).
v. Problem formulation: Reverse engineering causal models from multiple heterogeneous data sets.
vi. Modeling interventions/selection.
2. Logic-based causal discoveryi. Converting path constraints to logic formulae.
ii. Problem complexity.
iii. Conflict resolution.
iv. Existing algorithms.
v. Reasoning with logic based causal discovery.
vi. Non-trivial inferences-validation.
OUTLINE
41
![Page 42: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/42.jpg)
A
B
D
C E
42
LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERY
A
B
D
C E
A B C D E
Data Causal graph(s)(In)dependencies paths
…A B DC E
CA B D
Logic encoding Φ of path constraints in the causal graph
[E𝐴→D ∨ E𝐴→B ∧ E𝐵→D ∨E𝐴→C ∧ E𝐶→D ∨ ⋯ ]
⋮[E𝐴→C ∨ E𝐴→B ∧ E𝐵→C ∨
E𝐴↔C ∧ E𝐶→D ∨ … ]
Convert to logic formula!
Variables of the formula correspond to graph features (edges, orientations).
Truth setting assignments encode graphs that satisfy all path constraints after manipulation/selection.
![Page 43: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/43.jpg)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In an observational data set, Ind(A, C|∅)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given ∅ in 𝐺.
CONVERSION TO LOGIC FORMULA: EXAMPLE
43
B
CA
![Page 44: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/44.jpg)
• Edges of the graph as Boolean variables• E𝐴→𝐵 = 𝑡𝑟𝑢𝑒 if 𝐴 → 𝐵 in 𝐺, 𝑓𝑎𝑙𝑠𝑒 otherwise.
• E𝐴←𝐵 = 𝑡𝑟𝑢𝑒 if 𝐴 → 𝐵 in 𝐺, 𝑓𝑎𝑙𝑠𝑒 otherwise.
• E𝐴↔𝐵 = 𝑡𝑟𝑢𝑒 if 𝐴 ↔ 𝐵 in 𝐺, 𝑓𝑎𝑙𝑠𝑒 otherwise.
• E𝐴→𝐵 and E𝐴←𝐵 are mutually exclusive: ¬E𝐴→𝐵 ∨ ¬E𝐴←𝐵.
CONVERSION TO LOGIC FORMULA: EXAMPLE
44
B
CA
𝐸𝐴→𝐶 = 𝐹𝑎𝑙𝑠𝑒𝐸𝐴←𝐶 = 𝐹𝑎𝑙𝑠𝑒𝑬𝑨↔𝑪 = 𝑻𝒓𝒖𝒆
𝐸𝐴→𝐵 = 𝐹𝑎𝑙𝑠𝑒𝐸𝐴←𝐵 = 𝐹𝑎𝑙𝑠𝑒𝐸𝐴↔𝐵 = 𝐹𝑎𝑙𝑠𝑒
𝐸𝐵→𝐶 = 𝐹𝑎𝑙𝑠𝑒𝑬𝑩←𝑪 = 𝑻𝒓𝒖𝒆𝐸𝐵↔𝐶 = 𝐹𝑎𝑙𝑠𝑒
Assignments to Boolean variables correspond to graphs.
![Page 45: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/45.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE
45
B
CA
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In an observational data set, Ind(A, C|∅)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given ∅ in 𝐺.
![Page 46: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/46.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE
46
B
CA
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
A-B-C is not m-connecting
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In an observational data set, Ind(A, C|∅)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given ∅ in 𝐺.
![Page 47: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/47.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE
47
B
CA
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
A-B-C is not m-connecting¬(𝐸𝐴←𝐵∧ 𝐸𝐵←𝐶)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In an observational data set, Ind(A, C|∅)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given ∅ in 𝐺.
![Page 48: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/48.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE
48
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
B
CA
A-B-C is not m-connecting¬(𝐸𝐴←𝐵∧ 𝐸𝐵←𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵→𝐶)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In an observational data set, Ind(A, C|∅)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given ∅ in 𝐺.
![Page 49: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/49.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE
49
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
B
CA
A-B-C is not m-connecting¬(𝐸𝐴←𝐵∧ 𝐸𝐵←𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵↔𝐶)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In an observational data set, Ind(A, C|∅)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given ∅ in 𝐺.
![Page 50: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/50.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE
50
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
B
CA
A-B-C is not m-connecting¬(𝐸𝐴←𝐵∧ 𝐸𝐵←𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵↔𝐶)¬(𝐸𝐴→𝐵∧ 𝐸𝐵→𝐶)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In an observational data set, Ind(A, C|∅)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given ∅ in 𝐺.
![Page 51: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/51.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE
51
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
B
CA
A-B-C is not m-connecting¬(𝐸𝐴←𝐵∧ 𝐸𝐵←𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵↔𝐶)¬(𝐸𝐴→𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴↔𝐵∧ 𝐸𝐵→𝐶)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In an observational data set, Ind(A, C|∅)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given ∅ in 𝐺.
![Page 52: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/52.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE
52
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
B
CA
A-B-C is not m-connecting¬(𝐸𝐴←𝐵∧ 𝐸𝐵←𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵↔𝐶)¬(𝐸𝐴→𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴↔𝐵∧ 𝐸𝐵→𝐶)
Logic formula:¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶 ∧
¬(𝐸𝐴←𝐵∧ 𝐸𝐵←𝐶) ∧¬(𝐸𝐴←𝐵∧ 𝐸𝐵→𝐶) ∧¬(𝐸𝐴←𝐵∧ 𝐸𝐵↔𝐶) ∧¬(𝐸𝐴→𝐵∧ 𝐸𝐵→𝐶) ∧¬(𝐸𝐴↔𝐵∧ 𝐸𝐵→𝐶)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In an observational data set, Ind(A, C|∅)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given ∅ in 𝐺.
![Page 53: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/53.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE
53
B
CALogic formula:
¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶 ∧
¬(𝐸𝐴←𝐵∧ 𝐸𝐵←𝐶) ∧
¬(𝐸𝐴←𝐵∧ 𝐸𝐵→𝐶) ∧
¬(𝐸𝐴←𝐵∧ 𝐸𝐵↔𝐶) ∧
¬(𝐸𝐴→𝐵∧ 𝐸𝐵→𝐶) ∧
¬(𝐸𝐴↔𝐵∧ 𝐸𝐵→𝐶)
𝐸𝐴→𝐶 = 𝐹𝑎𝑙𝑠𝑒𝐸𝐴←𝐶 = 𝐹𝑎𝑙𝑠𝑒𝐸𝐴↔𝐶 = 𝐹𝑎𝑙𝑠𝑒
𝐸𝐴→𝐵 = 𝐹𝑎𝑙𝑠𝑒𝐸𝐴←𝐵 = 𝐹𝑎𝑙𝑠𝑒𝑬𝑨↔𝑩 = 𝑻𝒓𝒖𝒆
𝐸𝐵→𝐶 = 𝐹𝑎𝑙𝑠𝑒𝐸𝐵←𝐶 = 𝐹𝑎𝑙𝑠𝑒𝐸𝐵↔𝐶 = 𝐹𝑎𝑙𝑠𝑒
TRUE
![Page 54: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/54.jpg)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In a data set where 𝐵 is manipulated, Ind(A, C|∅)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given ∅ in 𝐺𝐵 .
CONVERSION TO LOGIC FORMULA: EXAMPLE (INTERVENTION)
54
B
CA
![Page 55: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/55.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE (INTERVENTION)
55
B
CA
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In a data set where 𝐵 is manipulated, Ind(A, C|∅)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given ∅ in 𝐺𝐵 .
![Page 56: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/56.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE (INTERVENTION)
56
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
B
CA
A-B-C is not m-connecting¬(𝐸𝐵→𝐴∧ 𝐸𝐵→𝐶)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In a data set where 𝐵 is manipulated, Ind(A, C|∅)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given ∅ in 𝐺𝐵 .
![Page 57: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/57.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE (INTERVENTION)
57
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
B
CA
A-B-C is not m-connecting¬(𝐸𝐵→𝐴∧ 𝐸𝐵→𝐶)
B has no incoming edges in 𝐺𝐵 .
Logic formula:¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶 ∧
¬(𝐸𝐴←𝐵∧ 𝐸𝐵→𝐶)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In a data set where 𝐵 is manipulated, Ind(A, C|∅)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given ∅ in 𝐺𝐵 .
![Page 58: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/58.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE
58
B
CALogic formula:
¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶 ∧
¬(𝐸𝐴←𝐵∧ 𝐸𝐵→𝐶) ∧
𝐸𝐴→𝐶 = 𝐹𝑎𝑙𝑠𝑒𝐸𝐴←𝐶 = 𝐹𝑎𝑙𝑠𝑒𝐸𝐴↔𝐶 = 𝐹𝑎𝑙𝑠𝑒
𝐸𝐴→𝐵 = 𝐹𝑎𝑙𝑠𝑒𝐸𝐴←𝐵 = 𝐹𝑎𝑙𝑠𝑒𝑬𝑨↔𝑩 = 𝑻𝒓𝒖𝒆
𝐸𝐵→𝐶 = 𝑇𝑟𝑢𝑒𝐸𝐵←𝐶 = 𝐹𝑎𝑙𝑠𝑒𝐸𝐵↔𝐶 = 𝐹𝑎𝑙𝑠𝑒
TRUE
𝐺
B
CA
𝐺𝐵
![Page 59: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/59.jpg)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In a data set selected based on 𝐵, Ind(A, C|𝑆 = 1)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given 𝑆 = 1 in 𝐺B.
CONVERSION TO LOGIC FORMULA: EXAMPLE (SELECTION)
59
B
CA
S=1
![Page 60: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/60.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE (SELECTION)
60
B
CA
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In a data set selected based on 𝐵, Ind(A, C|𝑆 = 1)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given 𝑆 = 1 in 𝐺B.
S=1
![Page 61: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/61.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE (SELECTION)
61
B
CA
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
A-B-C is not m-connecting¬(𝐸𝐴←𝐵∧ 𝐸𝐵←𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵↔𝐶)¬(𝐸𝐴→𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴↔𝐵∧ 𝐸𝐵→𝐶)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In a data set selected based on 𝐵, Ind(A, C|𝑆 = 1)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given 𝑆 = 1 in 𝐺B.
S=1
![Page 62: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/62.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE (SELECTION)
62
B
CA
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
A-B-C is not m-connecting¬(𝐸𝐴←𝐵∧ 𝐸𝐵←𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵↔𝐶)¬(𝐸𝐴→𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴↔𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴↔𝐵∧ 𝐸𝐵←𝐶)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In a data set selected based on 𝐵, Ind(A, C|𝑆 = 1)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given 𝑆 = 1 in 𝐺B.
S=1
![Page 63: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/63.jpg)
CONVERSION TO LOGIC FORMULA: EXAMPLE (SELECTION)
63
B
CA
A-C does not exist¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶
A-B-C is not m-connecting¬(𝐸𝐴←𝐵∧ 𝐸𝐵←𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴←𝐵∧ 𝐸𝐵↔𝐶)¬(𝐸𝐴→𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴↔𝐵∧ 𝐸𝐵→𝐶)¬(𝐸𝐴→𝐵∧ 𝐸𝐵←𝐶)¬(𝐸𝐴↔𝐵∧ 𝐸𝐵←𝐶)¬(𝐸𝐴→𝐵∧ 𝐸𝐵↔𝐶)¬(𝐸𝐴↔𝐵∧ 𝐸𝐵↔𝐶)
Logic formula:¬𝐸𝐴→𝐶 ∧ ¬𝐸𝐴←𝐶 ∧ ¬𝐸𝐴↔𝐶 ∧
¬(𝐸𝐴←𝐵∧ 𝐸𝐵←𝐶) ∧¬(𝐸𝐴←𝐵∧ 𝐸𝐵→𝐶) ∧¬(𝐸𝐴←𝐵∧ 𝐸𝐵↔𝐶) ∧¬(𝐸𝐴→𝐵∧ 𝐸𝐵→𝐶) ∧¬(𝐸𝐴↔𝐵∧ 𝐸𝐵→𝐶) ∧¬(𝐸𝐴→𝐵∧ 𝐸𝐵←𝐶) ∧¬(𝐸𝐴↔𝐵∧ 𝐸𝐵←𝐶) ∧¬(𝐸𝐴→𝐵∧ 𝐸𝐵↔𝐶) ∧¬(𝐸𝐴↔𝐵∧ 𝐸𝐵↔𝐶)
• Suppose you know nothing about the causal structure 𝐺 of 𝐴, 𝐵, 𝐶.
• In a data set selected based on 𝐵, Ind(A, C|𝑆 = 1)
• In path terms: ∄ m-connecting path between 𝐴 and 𝐶 given 𝑆 = 1 in 𝐺B.
S=1
![Page 64: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/64.jpg)
Path constraints corresponding to (conditional) dependencies and independencies from multiple datasets.
Information about the datasets Whether your samples were selected based on some variables.
Variables that were manipulated in your data set.
Many more ways to encode constraints into logic Different variable choices (e.g. edge *-*, orientations).
Different constraint choices depending on the problem at hand.
Ancestral paths
Inducing paths.
Colliders/non-colliders.
CONVERSION TO LOGIC FORMULA: INPUT CONSTRAINTS
64
![Page 65: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/65.jpg)
Logic-based causal discovery trivially and collectively handles cases for which no algorithm existed!
• Incorporating prior knowledge.• Algorithms for learning Bayesian networks can only enforce the presence/absence of direct edges.
• Easily impose presence/absence of direct edges, directed paths or m-connections (associations).
• root/leaf nodes.
• Learning semi-Markov causal graphs.• no learning algorithm until logic-based causal discovery.
•Combining heterogeneous data sets.
• Soft interventions.
• Sound and complete algorithms with incomplete knowledge (e.g. can not perform some tests of independence).
CONVERSION TO LOGIC FORMULA: VERSATILITY
65
![Page 66: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/66.jpg)
1. Integrative causal discoveryi. Motivation.
ii. Causal models.
iii. m-separation.
iv. Reverse engineering causal models (single data set).
v. Problem formulation: Reverse engineering causal models from multiple heterogeneous data sets.
vi. Modeling interventions/selection.
2. Logic-based causal discoveryi. Converting path constraints to logic formulae.
ii. Problem complexity.
iii. Conflict resolution.
iv. Existing algorithms.
v. Reasoning with logic based causal discovery.
vi. Non-trivial inferences-validation.
OUTLINE
66
![Page 67: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/67.jpg)
A
B
D
C E
67
LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERY
A
B
D
C E
A B C D E
Data Causal graph(s)(In)dependencies Paths
…A B DC E
CA B D
[E𝐴→D ∨ E𝐴→B ∧ E𝐵→D ∨E𝐴→C ∧ E𝐶→D ∨
⋮[E𝐴→C ∨ E𝐴→B ∧ E𝐵→C ∨
E𝐴↔C ∧ E𝐶→D
Logic formula
Exponential number of 1.Independencies
2.Paths3. Solutions
![Page 68: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/68.jpg)
For a data set with 10 variables:
28 = 256 different conditioning sets
For each conditioning set, you need to consider all possible paths with up to 9 edges:
𝑘=210 8
10−𝑘= 1435 paths per pair of variables.
In total: 102
= 45 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒 𝑝𝑎𝑖𝑟𝑠 × 256 𝑐𝑜𝑛𝑑 𝑠𝑒𝑡𝑠 × 1435 𝑝𝑎𝑡ℎ𝑠 = 16531200 𝑝𝑎𝑡ℎ 𝑐𝑜𝑛𝑠𝑡𝑟𝑎𝑖𝑛𝑡𝑠.
For a network of 10 variables:
135 possible edges.
2135~ 1040 different graphs.
PROBLEM COMPLEXITY: EXAMPLE
68
Brute force approach only works for ~10 variables regardless of encoding.
Several heuristics for scaling up (depending on the algorithm).
You can take into account all dependencies and independencies, even for a small number of variables.
![Page 69: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/69.jpg)
A
B
D
C E
69
LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERY
A
B
D
C E
A B C D E
Data Causal graph(s)(In)dependencies Paths
…A B DC E
CA B D
[E𝐴→D ∨ E𝐴→B ∧ E𝐵→D ∨E𝐴→C ∧ E𝐶→D ∨
⋮[E𝐴→C ∨ E𝐴→B ∧ E𝐵→C ∨
E𝐴↔C ∧ E𝐶→D ∨
Logic formula
Reduce the number of independencies:
Run FCI and use only the tests performed by FCI.
Limit max conditioning set size.
![Page 70: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/70.jpg)
A
B
D
C E
70
LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERY
A
B
D
C E
A B C D E
Data Causal graph(s)(In)dependencies Paths
…A B DC E
CA B D
[E𝐴→D ∨ E𝐴→B ∧ E𝐵→D ∨E𝐴→C ∧ E𝐶→D ∨
⋮[E𝐴→C ∨ E𝐴→B ∧ E𝐵→C ∨
E𝐴↔C ∧ E𝐶→D ∨
Logic formula
Reduce the number of paths:
Use inducing paths that connect paths on the graph to ∃ of independence (given any set).
Limit the maximum path length.
![Page 71: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/71.jpg)
A
B
D
C E
71
LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERY
A
B
D
C E
A B C D E
Data Causal graph(s)(In)dependencies Paths
…A B DC E
CA B D
[E𝐴→D ∨ E𝐴→B ∧ E𝐵→D ∨E𝐴→C ∧ E𝐶→D ∨
⋮[E𝐴→C ∨ E𝐴→B ∧ E𝐵→C ∨
E𝐴↔C ∧ E𝐶→D ∨
Logic formula
Need a clever way to encode constraints!
e.g. recursively encode paths.
Convert to CNF for most SAT solvers.
![Page 72: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/72.jpg)
A
B
D
C E
72
LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERY
A
B
D
C E
A B C D E
Data Causal graph(s)(In)dependencies Paths
…A B DC E
CA B D
[E𝐴→D ∨ E𝐴→B ∧ E𝐵→D ∨E𝐴→C ∧ E𝐶→D ∨
⋮[E𝐴→C ∨ E𝐴→B ∧ E𝐵→C ∨
E𝐴↔C ∧ E𝐶→D ∨
Logic formula
No need to enumerate all solutions!
Query the formula for• A single causal graph.• A causal graph with specific
features.• Features that are invariant
in all possible causal graphs.
![Page 73: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/73.jpg)
Breast Cancer
Protein C
ContraceptivesThrombosis
Protein Z
Protein E
73
SUMMARIZING PAIRWISE RELATIONS
Absent edges: Absent in allsolutions
![Page 74: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/74.jpg)
Breast Cancer
Protein C
ContraceptivesThrombosis
Protein Z
Protein E
74
SUMMARIZING PAIRWISE RELATIONS
solid edges: present in allsolutions
Absent edges: Absent in allsolutions
![Page 75: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/75.jpg)
Breast Cancer
Protein C
ContraceptivesThrombosis
Protein Z
Protein E
75
SUMMARIZING PAIRWISE RELATIONS
dashed edges: present in somesolutions
solid edges: present in allsolutions
Absent edges: Absent in allsolutions
![Page 76: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/76.jpg)
Breast Cancer
Protein C
ContraceptivesThrombosis
Protein Z
Protein E
76
SUMMARIZING PAIRWISE RELATIONS
dashed edges: present in somesolutions
solid edges: present in allsolutions
solid endpoints: same orientation in all solutions
Absent edges: Absent in allsolutions
![Page 77: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/77.jpg)
Breast Cancer
Protein C
ContraceptivesThrombosis
Protein Z
Protein E
77
SUMMARIZING PAIRWISE RELATIONS
dashed edges: present in somesolutions
solid edges: present in allsolutions
solid endpoints: same orientation in all solutions
Circle endpoints: orientation varies in different solutions
Absent edges: Absent in allsolutions
![Page 78: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/78.jpg)
1. Integrative causal discoveryi. Motivation.
ii. Causal models.
iii. m-separation.
iv. Reverse engineering causal models (single data set).
v. Problem formulation: Reverse engineering causal models from multiple heterogeneous data sets.
vi. Modeling interventions/selection.
2. Logic-based causal discoveryi. Converting path constraints to logic formulae.
ii. Problem complexity.
iii. Conflict resolution.
iv. Existing algorithms.
v. Reasoning with logic based causal discovery.
vi. Non-trivial inferences-validation.
OUTLINE
78
![Page 79: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/79.jpg)
A
B
D
C E
79
STATISTICAL ERRORS RESULT IN CONFLICTING INPUTS
A
B
D
C E
A B C D E
Data Causal graph(s)(In)dependencies Paths
…A B DC E
CA B D
[E𝐴→D ∨ E𝐴→B ∧ E𝐵→D ∨E𝐴→C ∧ E𝐶→D ∨
⋮[E𝐴→C ∨ E𝐴→B ∧ E𝐵→C ∨
E𝐴↔C ∧ E𝐶→D ∨
Logic formulaStatistical errors
Conflicting constraints
Unsatisfiableformula
![Page 80: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/80.jpg)
80
TESTING (CONDITIONAL) INDEPENDENCE
𝑝= 3.2*10−5
0
1
𝑝= 0.54
Threshold Independence
Dependence
p-value: 𝑃(𝑫|𝐼𝑛𝑑) (VERY loose interpretation)
![Page 81: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/81.jpg)
81
TESTING (CONDITIONAL) INDEPENDENCE
0
1
Threshold Independence
Dependence
𝐷1: 𝐼𝑛𝑑 𝐴, 𝐷 ∅
¬𝑚 − 𝑐𝑜𝑛 𝐴, 𝐷 ∅
𝐷2: 𝐷𝑒𝑝 𝐴, 𝐷 ∅
𝑚 − 𝑐𝑜𝑛 𝐴, 𝐷 ∅
𝑝(𝐴,𝐷|∅)𝐷1= 0.54
𝑝(𝐴,𝐷|∅)𝐷2= 3.2*10−5
What happens with statistical errors?
Conflicts make SAT instance unsatisfiable!
Different observational data sets, same relationship, different p-values.
p-value: 𝑃(𝑫|𝐼𝑛𝑑) (VERY loose interpretation)
![Page 82: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/82.jpg)
82
TESTING (CONDITIONAL) INDEPENDENCE
0
1
Threshold Independence
Dependence
𝑝(𝐴,𝐷|∅)𝐷1= 0.54
𝑝(𝐴,𝐷|∅)𝐷2= 3.2*10−5
What is more probable: a
dependence with p-value 3.2*10−5 or
independence with p-value 0.54?
You need to estimate 𝑃 𝐼𝑛𝑑 𝑫
How can you decide if Independence is more probable than dependence?
Different observational data sets, same relationship, different p-values.
𝐷1: 𝐼𝑛𝑑 𝐴, 𝐷 ∅
¬𝑚 − 𝑐𝑜𝑛 𝐴, 𝐷 ∅
𝐷2: 𝐷𝑒𝑝 𝐴, 𝐷 ∅
𝑚 − 𝑐𝑜𝑛 𝐴, 𝐷 ∅
p-value: 𝑃(𝑫|𝐼𝑛𝑑) (VERY loose interpretation)
![Page 83: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/83.jpg)
83
ESTIMATING 𝑃 𝐼𝑛𝑑 𝑫 USING BAYESIAN SCORING (1)
• You want to estimate 𝑃(𝐼𝑛𝑑(𝑋, 𝑌|𝑍)|𝑫)
• Score every possible DAG over 𝑋, 𝑌, 𝑍: 𝑃(𝑫|𝐺).• You can use BDE, BGE to compute 𝑃 𝑫 𝐺 .
• Some of these DAGs entail dependence (m-connection, some independence (m-separation).• Define a prior over graphs.
• Take the weighted average:• 𝑃 𝐼𝑛𝑑(𝑋, 𝑌|𝒁) 𝑫 ∝ 𝐺:𝐺 𝑒𝑛𝑡𝑎𝑖𝑙𝑠 𝐼𝑛𝑑(𝑋,𝑌|𝒁) 𝑃 𝑫 𝐺 × 𝑃(𝐺)
• Exponential number of DAGs.• Use one graph per Markov equivalence class (still exponential).• Still not possible for more than 5-6 variables.
X Y
Z
⋮
X Y
Z𝐺1:
𝐺𝑛
Ind(X, Y|Z)
X Y
Z𝐺2:
X Y
Z𝐺𝑖:
⋮
Ind(X, Y|Z)
Dep(X, Y|Z)
Dep(X, Y|Z)
[BCCD, Claassen and Heskes, UAI 2012]
![Page 84: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/84.jpg)
84
ESTIMATING 𝑃 𝐼𝑛𝑑 𝑫 USING BAYESIAN SCORING (2)
• You want to estimate 𝑃(𝐼𝑛𝑑(𝑋, 𝑌|𝒁)|𝑫)
• Independence 𝐼𝑛𝑑 𝑋, 𝑌 𝒁 : 𝑃 𝑋, 𝑌 𝒁 = 𝑃 𝑋 𝒁 𝑃(𝑌|𝒁)• Dependence 𝐷𝑒𝑝 𝑋, 𝑌 𝒁 : 𝑃 𝑋, 𝑌 𝒁 = 𝑃 𝑋 𝒁 𝑃(𝑌|𝑋, 𝒁)
• 𝑃 𝐼𝑛𝑑(𝑋, 𝑌|𝒁) 𝑫 =𝑃(𝑌|𝒁) 𝜋0
𝑃(𝑌|𝒁)𝜋0+𝑃(𝑌|𝑋,𝒁)(1−𝜋0).
• Use BDE, BGE to estimate 𝑃(𝑌|𝒁), 𝑃(𝑌|𝑋, 𝒁).• 𝜋0: Prior for independence is an input parameter.
𝑋 𝑍𝑛
𝑌
𝑍1
𝑋 𝑍𝑛
𝑌
𝑍1
vs.
[M&B, Margaritis and Bromberg, CI 2009]
𝑃(𝑌|𝒁)
𝑃(𝑌|𝑋, 𝒁)
![Page 85: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/85.jpg)
• p-values coming from independence follow a 𝐵𝑒𝑡𝑎(1, 1) distribution• p-values coming from dependence follow a distribution in (0, 1) with declining
density• Can be modeled with a 𝐵𝑒𝑡𝑎 𝜉, 1 , 𝜉 ∈ (0, 1) distribution.
85
ESTIMATING 𝑃 𝐼𝑛𝑑 𝑫 FROM P-VALUES
[PROPER, Triantafillou et al, PGM 2014]
![Page 86: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/86.jpg)
• p-values coming from independence follow a 𝐵𝑒𝑡𝑎(1, 1) distribution• p-values coming from dependence follow a distribution in (0, 1) with declining
density• Can be modeled with a 𝐵𝑒𝑡𝑎 𝜉, 1 , 𝜉 ∈ (0, 1) distribution.
• Let 𝜋0 be the proportion of independencies.
• 𝑓 𝑝 𝜋𝑜, 𝜉 = 𝜋0 + 1 − 𝜋0 𝜉𝑝𝜉−1.
86
ESTIMATING 𝑃 𝐼𝑛𝑑 𝑫 FROM P-VALUES
[PROPER, Triantafillou et al, PGM 2014]
![Page 87: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/87.jpg)
• p-values coming from independence follow a 𝐵𝑒𝑡𝑎(1, 1) distribution• p-values coming from dependence follow a distribution in (0, 1) with declining
density• Can be modeled with a 𝐵𝑒𝑡𝑎 𝜉, 1 , 𝜉 ∈ (0, 1) distribution.
• Let 𝜋0 be the proportion of independencies.
• 𝑓 𝑝 𝜋𝑜, 𝜉 = 𝜋0 + 1 − 𝜋0 𝜉𝑝𝜉−1.
• You can find estimate 𝜋0, 𝜉 from the empirical distribution of your p-values• Find 𝜋0 using [Storey and Tibshirani, 2003] (assumes i.i.d. p-values)
• Find 𝜉 by minimizing negative log likelihood
87
ESTIMATING 𝑃 𝐼𝑛𝑑 𝑫 FROM P-VALUES
[PROPER, Triantafillou et al, PGM 2014]
![Page 88: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/88.jpg)
• p-values coming from independence follow a 𝐵𝑒𝑡𝑎(1, 1) distribution• p-values coming from dependence follow a distribution in (0, 1) with declining
density• Can be modeled with a 𝐵𝑒𝑡𝑎 𝜉, 1 , 𝜉 ∈ (0, 1) distribution.
• Let 𝜋0 be the proportion of independencies.
• 𝑓 𝑝 𝜋𝑜, 𝜉 = 𝜋0 + 1 − 𝜋0 𝜉𝑝𝜉−1.
• You can find estimate 𝜋0, 𝜉 from the empirical distribution of your p-values• Find 𝜋0 using [Storey and Tibshirani, 2003] (assumes i.i.d. p-values)
• Find 𝜉 by minimizing negative log likelihood
88
ESTIMATING 𝑃 𝐼𝑛𝑑 𝑫 FROM P-VALUES
𝑃(𝐼𝑛𝑑|𝑝) =
𝜋0
1 − 𝜋0 𝜉𝑝(1− 𝜉)
1 + 𝜋0
1 − 𝜋0 𝜉𝑝(1− 𝜉) [PROPER, Triantafillou et al, PGM 2014]
![Page 89: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/89.jpg)
• Bayesian methods• Use the data directly.
• No problem if you have data sets with different sample sizes etc.
• Computationally expensive.
• Choose a prior for 𝜋0.
•PROPER (based on p-values)• Scalable, no computational overhead, benefits from larger p-value populations (more tests).
• Estimate 𝜋0 from the data.
• p-values are not i.i.d.
ESTIMATING 𝑃 𝐼𝑛𝑑 𝑫
89
![Page 90: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/90.jpg)
90
CONFLICT RESOLUTION STRATEGIES
P(constraint) Ind/Dep path constraint
0.999 Dep ∃ m-connecting path from A to D given ∅ in 𝑆𝐼𝑛
0.998 Ind ∄ m-connecting path from A to D given ∅ in 𝑆𝐼1
⋮ ⋮ ⋮
0.510 Dep ∃ m-connecting path from A to B given ∅ in 𝑆𝐼1
• Assign weights according to P(constraint), maximize the sum of weights.
• Rank by probability, greedily satisfy constraints.
Maximizing sum of weights is the best strategyUse greedy to scale up.
![Page 91: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/91.jpg)
1. Integrative causal discoveryi. Motivation.
ii. Causal models.
iii. m-separation.
iv. Reverse engineering causal models (single data set).
v. Problem formulation: Reverse engineering causal models from multiple heterogeneous data sets.
vi. Modeling interventions/selection.
2. Logic-based causal discoveryi. Converting path constraints to logic formulae.
ii. Problem complexity.
iii. Conflict resolution.
iv. Existing algorithms.
v. Reasoning with logic based causal discovery.
vi. Non-trivial inferences-validation.
OUTLINE
91
![Page 92: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/92.jpg)
Vary in: Type of constraints:
different types of paths (m-connecting, inducing, ancestral). translation to logic formula.
Types of heterogeneity: Soft/hard interventions, selection.
Preprocessing: Heuristics to limit number of constraints / paths.
Conflict Resolution Method for calculating probabilities. Conflict resolution strategy (greedy/ max SAT / weighted max SAT).
CS solver Initially SAT solvers, more recently ASP.
Scalability Depends on choices above. Be exact/ focus on scalability. Difficult to determine huge variance depending on the problem.
EXISTING ALGORITHMS
92
Implementations varyheuristics are typically easy to incorporate in any algorithm.-maximum conditioning set size/ path length.-greedy /weighted max SAT.
![Page 93: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/93.jpg)
CSAT+
93
Type of Constraints m-connecting paths, inducing paths.
Type of Heterogeneity
Overlapping variables.
Preprocessing Runs FCI on multiple data sets.Additional preprocessing rules for additional edge removals/orientations.
Conflict resolution None (oracle only)
CSP solver MINISAT
Scalability ~37 variables (ALARM network)
[Triantafillou, Tsamardinos and Tollis, AISTATS 2010]
![Page 94: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/94.jpg)
LOCI
94
[Claassen and Heskes, UAI 2011]
Type of Constraints ancestral paths.Converts [minimal] conditional independencies to ancestral relations:𝐼𝑛𝑑 𝑋, 𝑌 𝒁 ⇒ 𝒁 → ⋯ → 𝑋 ∨ 𝒁 → ⋯ → 𝑌
Type of Heterogeneity
None (substitutes FCI orientation steps).
Preprocessing FCI skeleton step.
Conflict resolution None (single data set, runs similar to FCI orientation rules)
CSP solver custom set of rules
Scalability unknown (probably similar to FCI).
![Page 95: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/95.jpg)
SAT-BASED CAUSAL DISCOVERY
95
Type of Constraints m-connecting paths
Type of Heterogeneity
Overlapping variables, interventions also allows cycles.
Preprocessing None.Can use a subset of (in) dependencies depending on assumptions (e.g. FCI tests only)
Conflict resolution None (oracle only)
CSP solver MINISAT
Scalability 8-12 variables
[Hyttinen, Hoyer, Eberhardt and Järvisalo, UAI 2013]
![Page 96: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/96.jpg)
96
CONSTRAINT-BASED CAUSAL DISCOVERY
Type of Constraints m-connecting paths.encoded in ASP based on marginalization and conditioning.
Type of Heterogeneity
Overlapping variables, interventions allows cycles
Preprocessing none
Conflict resolution Default: based on M&B, maximize sum of weights (find global optimum),also tried maximizing the number of independencies/ number of constraints
CSP solver ASP
Scalability 7 variables
[Hyttinen, Eberhardt and Järvisalo, UAI 2014]
![Page 97: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/97.jpg)
COMBINE
97
Type of Constraints inducing pathsDrastically reduces the number of constraints (∃, ∄ path) to 1 per variable pair & data set(compared to 2𝑛)
Type of Heterogeneity
Overlapping variables, interventions
Preprocessing FCI on each data set.
Conflict resolution Default: based on PROPER, greedy search. also implemented: BCCD , weighted maxSAT.
CSP solver MINISAT
Scalability 100 variables (additionally limits maximum path length)
[Triantafillou and Tsamardinos, JMLR 2015]
![Page 98: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/98.jpg)
ETIO
98
[Borboudakis and Tsamardinos, KDD 2016]
Type of Constraints m-connecting paths.encoded in ASP based on extension of the Bayes-Ball algorithm (used to determine m-connections/m-separations in graphs) for SMCGs with selection.
Type of Heterogeneity
Overlapping variables, interventions, selection.
Preprocessing none
Conflict resolution based on PROPER/M&B, greedy
CSP solver ASP
Scalability 10-15 variables
![Page 99: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/99.jpg)
ACI
99
[S. Magliacane, T. Claassen, J.M. Mooij, arXiv]
Type of Constraints m-connections, ancestry relations
Type of Heterogeneity
Overlapping variables, various types of interventions
Preprocessing none
Conflict resolution based on M&B, weighted maxSAT
CSP solver ASP
Scalability 10-15 variables
![Page 100: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/100.jpg)
100
MORE
• Using conversion to logic to incorporate prior knowledge in maximal ancestral graphs.• [Borboudakis, Triantafillou and Tsamardinos, ESANN 2011].
• Using conversion to logic for causal discovery from time-course data• Causal Discovery from Subsampled Time Series Data by Constraint
Optimization, [Hyttinen, Plis, Järvisalo, Eberhardt and Danks, arXiv, 2016]
• Using conversion to logic for identifying chain graphs.• Learning Optimal Chain Graphs with Answer Set Programming[Sonntag,
Järvisalo, Penã, Hyttinen, UAI 2015]
• Using conversion to logic to identify semi-Markov causal graphs.• [Penã, UAI 2016]
![Page 101: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/101.jpg)
Different data distributions, same causal mechanism: use causal modeling to connect.
Algorithms can handle datasets of different variable sets, different experimental conditions, prior causal knowledge.
Identify the set of causal graphs that simultaneously fit all datasets .
Convert problem to SAT or ASP.
Logic formula encodes a set of causal models that simultaneously fit all the data sets.
OVERVIEW
101
![Page 102: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/102.jpg)
-How can you reason with this set of models?
-Is it useful? Do you make additional inferences than analyzing each data set in isolation?
QUESTIONS
102
![Page 103: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/103.jpg)
1. Integrative causal discoveryi. Motivation.
ii. Causal models.
iii. m-separation.
iv. Reverse engineering causal models (single data set).
v. Problem formulation: Reverse engineering causal models from multiple heterogeneous data sets.
vi. Modeling interventions/selection.
2. Logic-based causal discoveryi. Converting path constraints to logic formulae.
ii. Problem complexity.
iii. Conflict resolution.
iv. Existing algorithms.
v. Reasoning with logic based causal discovery.
vi. Non-trivial inferences-validation.
OUTLINE
103
![Page 104: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/104.jpg)
104
ESTIMATING CAUSAL EFFECTS
B
CA𝐶 = 𝑓 𝐴
𝐵 = 𝑓 𝐶
You are interested in computing 𝑃(𝐵|𝑑𝑜(𝐴 = 𝑎))In general, 𝑃 𝐵 𝑑𝑜 𝐴 = 𝑎 ≠ 𝑃 𝐵 𝐴
If you know the causal graph, you can use the rules of do-calculus to transform post-intervention probabilities to pre-intervention probabilities.
![Page 105: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/105.jpg)
105
ESTIMATING CAUSAL EFFECTS
B
CA𝐶 = 𝑓 𝐴
𝐵 = 𝑓 𝐶
You are interested in computing 𝑃(𝐵|𝑑𝑜(𝐴 = 𝑎))In general, 𝑃 𝐵 𝑑𝑜 𝐴 = 𝑎 ≠ 𝑃 𝐵 𝐴
If you know the causal graph, you can use the rules of do-calculus to transform post-intervention probabilities to pre-intervention probabilities.
[Rule 1] 𝐼𝑛𝑑 𝑌, 𝑍 𝑋, 𝑊 𝐺𝑋 ⇒ 𝑃 𝑦 𝑑𝑜 𝑥 , 𝑧, 𝑤 = 𝑃(𝑦|𝑑𝑜 𝑥 , 𝑤).
[Rule 2] 𝐼𝑛𝑑 𝑌, 𝐼𝑍 𝑋, 𝑍, 𝑊 𝐺𝑋 ⇒ 𝑃 𝑦 𝑑𝑜 𝑥 , 𝑑𝑜(𝑧), 𝑤 = 𝑃 𝑦 𝑑𝑜 𝑥 , 𝑧, 𝑤 .
[Rule 3] 𝐼𝑛𝑑 𝑌, 𝐼𝑍 𝑋, 𝑊 𝐺𝑋 ⇒ 𝑃 𝑦 𝑑𝑜 𝑥 , 𝑑𝑜(𝑧), 𝑤 = 𝑃 𝑦 𝑑𝑜 𝑥 , 𝑤 .
Insert/delete observations
Exchange action/observation
Insert/delete action
Check m-separations ⇒Apply rules until you have a formula with pre-intervention probabilities
[Shpitser and Pearl (2006): Return a formula if identifiable]
![Page 106: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/106.jpg)
106
DO-CALCULUS WHEN THE GRAPH IS UNKNOWN
[Hyttinen, Eberhardt and Järvisalo, UAI 2015]
Constraints in logic formula ΦCausal effect 𝑃 𝐵 𝑑𝑜 𝐴
∃ m-connecting path from A to B given ∅∃ m-connecting path from A to B given 𝐶∃ m-connecting path from A to C given ∅∃ m-connecting path from A to C given 𝐵∃ m-connecting path from B to C given ∅∃ m-connecting path from B to C given 𝐴∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝐴 to C∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 A to B∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 B to C
![Page 107: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/107.jpg)
B
C A
107
DO-CALCULUS WHEN THE GRAPH IS UNKNOWN
[Hyttinen, Eberhardt and Järvisalo, UAI 2015]
∃ m-connecting path from A to B given ∅∃ m-connecting path from A to B given 𝐶∃ m-connecting path from A to C given ∅∃ m-connecting path from A to C given 𝐵∃ m-connecting path from B to C given ∅∃ m-connecting path from B to C given 𝐴∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝐴 to C∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 A to B∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 B to C
Constraints in logic formula ΦFind a graph
consistent with Φ Causal effect 𝑃 𝐵 𝑑𝑜 𝐴
![Page 108: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/108.jpg)
B
C A
108
DO-CALCULUS WHEN THE GRAPH IS UNKNOWN
[Hyttinen, Eberhardt and Järvisalo, UAI 2015]
∃ m-connecting path from A to B given ∅∃ m-connecting path from A to B given 𝐶∃ m-connecting path from A to C given ∅∃ m-connecting path from A to C given 𝐵∃ m-connecting path from B to C given ∅∃ m-connecting path from B to C given 𝐴∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝐴 to C∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 A to B∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 B to C
Constraints in logic formula ΦFind a graph
consistent with Φ Causal effect 𝑃 𝐵 𝑑𝑜 𝐴Shpitser and Pearl (2006)
𝐹1 =
𝑐
𝑃 𝑏 𝑎, 𝑐 𝑃(𝑐)
![Page 109: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/109.jpg)
B
C A
109
DO-CALCULUS WHEN THE GRAPH IS UNKNOWN
[Hyttinen, Eberhardt and Järvisalo, UAI 2015]
∃ m-connecting path from A to B given ∅∃ m-connecting path from A to B given 𝐶∃ m-connecting path from A to C given ∅∃ m-connecting path from A to C given 𝐵∃ m-connecting path from B to C given ∅∃ m-connecting path from B to C given 𝐴∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝐴 to C∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 A to B∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 B to C(∃ m-connecting path from 𝐼𝐴 to C given ∅ ∨∃ m-connecting path from 𝐼𝐴 to B given 𝐴, 𝐶)
Constraints in logic formula ΦFind a graph
consistent with Φ Causal effect 𝑃 𝐵 𝑑𝑜 𝐴
Identify and negate the m-separations used to derive the formula
Shpitser and Pearl (2006)
𝐹1 =
𝑐
𝑃 𝑏 𝑎, 𝑐 𝑃(𝑐)
![Page 110: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/110.jpg)
B
C A
110
DO-CALCULUS WHEN THE GRAPH IS UNKNOWN
[Hyttinen, Eberhardt and Järvisalo, UAI 2015]
∃ m-connecting path from A to B given ∅∃ m-connecting path from A to B given 𝐶∃ m-connecting path from A to C given ∅∃ m-connecting path from A to C given 𝐵∃ m-connecting path from B to C given ∅∃ m-connecting path from B to C given 𝐴∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝐴 to C∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 A to B∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 B to C(∃ m-connecting path from 𝐼𝐴 to C given ∅ ∨∃ m-connecting path from 𝐼𝐴 to B given 𝐴, 𝐶)
Constraints in logic formula ΦFind a graph
consistent with Φ Causal effect 𝑃 𝐵 𝑑𝑜 𝐴Shpitser and Pearl (2006)
𝐹1 =
𝑐
𝑃 𝑏 𝑎, 𝑐 𝑃(𝑐)
B
C A
![Page 111: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/111.jpg)
B
C A
111
DO-CALCULUS WHEN THE GRAPH IS UNKNOWN
[Hyttinen, Eberhardt and Järvisalo, UAI 2015]
∃ m-connecting path from A to B given ∅∃ m-connecting path from A to B given 𝐶∃ m-connecting path from A to C given ∅∃ m-connecting path from A to C given 𝐵∃ m-connecting path from B to C given ∅∃ m-connecting path from B to C given 𝐴∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝐴 to C∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 A to B∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 B to C(∃ m-connecting path from 𝐼𝐴 to C given ∅ ∨∃ m-connecting path from 𝐼𝐴 to B given 𝐴, 𝐶)
Constraints in logic formula ΦFind a graph
consistent with Φ Causal effect 𝑃 𝐵 𝑑𝑜 𝐴Shpitser and Pearl (2006)
𝐹1 =
𝑐
𝑃 𝑏 𝑎, 𝑐 𝑃(𝑐)
B
C A𝐹2 = 𝑃(𝑏)
![Page 112: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/112.jpg)
B
C A
112
DO-CALCULUS WHEN THE GRAPH IS UNKNOWN
[Hyttinen, Eberhardt and Jarvislao, UAI 2015]
∃ m-connecting path from A to B given ∅∃ m-connecting path from A to B given 𝐶∃ m-connecting path from A to C given ∅∃ m-connecting path from A to C given 𝐵∃ m-connecting path from B to C given ∅∃ m-connecting path from B to C given 𝐴∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝐴 to C∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 A to B∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 B to C(∃ m-connecting path from 𝐼𝐴 to C given ∅ ∨∃ m-connecting path from 𝐼𝐴 to B given 𝐴, 𝐶)∃ m-connecting path from 𝐼𝐴 to B given ∅)
Constraints in logic formula ΦFind a graph
consistent with Φ Causal effect 𝑃 𝐵 𝑑𝑜 𝐴Shpitser and Pearl (2006)
𝐹1 =
𝑐
𝑃 𝑏 𝑎, 𝑐 𝑃(𝑐)
B
C A
Identify and negate the m-separations used to derive the formula
𝐹2 = 𝑃(𝑏)
![Page 113: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/113.jpg)
B
C A
113
DO-CALCULUS WHEN THE GRAPH IS UNKNOWN
[Hyttinen, Eberhardt and Järvisalo, UAI 2015]
∃ m-connecting path from A to B given ∅∃ m-connecting path from A to B given 𝐶∃ m-connecting path from A to C given ∅∃ m-connecting path from A to C given 𝐵∃ m-connecting path from B to C given ∅∃ m-connecting path from B to C given 𝐴∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝐴 to C∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 A to B∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 B to C(∃ m-connecting path from 𝐼𝐴 to C given ∅ ∨∃ m-connecting path from 𝐼𝐴 to B given 𝐴, 𝐶)∃ m-connecting path from 𝐼𝐴 to B given ∅)
Constraints in logic formula ΦFind a graph
consistent with Φ Causal effect 𝑃 𝐵 𝑑𝑜 𝐴Shpitser and Pearl (2006)
𝐹1 =
𝑐
𝑃 𝑏 𝑎, 𝑐 𝑃(𝑐)
B
C A𝐹2 = 𝑃(𝑏)
B
C A𝑁𝐴
.
.
.
UNSAT
![Page 114: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/114.jpg)
B
C A
114
DO-CALCULUS WHEN THE GRAPH IS UNKNOWN
[Hyttinen, Eberhardt and Järvisalo, UAI 2015]
∃ m-connecting path from A to B given ∅∃ m-connecting path from A to B given 𝐶∃ m-connecting path from A to C given ∅∃ m-connecting path from A to C given 𝐵∃ m-connecting path from B to C given ∅∃ m-connecting path from B to C given 𝐴∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝐴 to C∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 A to B∄ 𝑑𝑖𝑟𝑒𝑐𝑡𝑒𝑑 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 B to C(∃ m-connecting path from 𝐼𝐴 to C given ∅ ∨∃ m-connecting path from 𝐼𝐴 to B given 𝐴, 𝐶)∃ m-connecting path from 𝐼𝐴 to B given ∅)
Constraints in logic formula ΦFind a graph
consistent with Φ Causal effect 𝑃 𝐵 𝑑𝑜 𝐴Shpitser and Pearl (2006)
𝐹1 =
𝑐
𝑃 𝑏 𝑎, 𝑐 𝑃(𝑐)
B
C A𝐹2 = 𝑃(𝑏)
B
C A𝑁𝐴
.
.
.
UNSAT
![Page 115: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/115.jpg)
-How can you reason with this set of models?You can use do-calculus and estimate (a population of) causal effects.
-Is it useful? Do you make additional inferences than analyzing each data set in isolation?
QUESTIONS
115
![Page 116: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/116.jpg)
1. Integrative Causal Discoveryi. Motivation.
ii. Causal models.
iii. m-separation.
iv. Reverse engineering causal models (single data set).
v. Problem formulation: Reverse engineering causal models from multiple heterogeneous data sets.
vi. Modeling interventions/selection.
2. Logic-based causal discoveryi. Converting path constraints to logic formulae.
ii. Problem complexity.
iii. Conflict resolution.
iv. Existing algorithms.
v. Reasoning with logic based causal discovery.
vi. Non-trivial inferences-validation.
OUTLINE
116
![Page 117: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/117.jpg)
117
EXAMPLE INPUT- OUTPUT
Y X W𝑋, 𝑊|𝑌 Ind
𝑋, 𝑊|∅ Dep
𝑋, 𝑌|∅ Dep
𝑌, 𝑊|∅ Dep
𝑌, 𝑋|𝑊 Dep
𝑌, 𝑊|𝑋 Dep
X W Z
Common variables
𝑋, 𝑊|𝑍 Ind
𝑋, 𝑊|∅ Dep
𝑋, 𝑍|∅ Dep
𝑍, 𝑊|∅ Dep
𝑍, 𝑊|𝑋 Dep
𝑍, 𝑋|𝑊 Dep
X
Y
Z
W
∄𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 𝑌 ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑌 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑌 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑌 𝑔𝑖𝑣𝑒𝑛 𝑊 ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑌 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 𝑋 ∧
∄𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 𝑍 ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑍 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑍𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑍 𝑔𝑖𝑣𝑒𝑛 𝑊 ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑍 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 𝑋
Formula ΦCI pattern 𝐶1
CI pattern 𝐶2
Dataset 𝐷2
Dataset 𝐷1
Summary of solutions
![Page 118: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/118.jpg)
118
EXAMPLE INPUT- OUTPUT
Y X W𝑋, 𝑊|𝑌 Ind
𝑋, 𝑊|∅ Dep
𝑋, 𝑌|∅ Dep
𝑌, 𝑊|∅ Dep
𝑌, 𝑋|𝑊 Dep
𝑌, 𝑊|𝑋 Dep
X W Z
Common variables
𝑋, 𝑊|𝑍 Ind
𝑋, 𝑊|∅ Dep
𝑋, 𝑍|∅ Dep
𝑍, 𝑊|∅ Dep
𝑍, 𝑊|𝑋 Dep
𝑍, 𝑋|𝑊 Dep
X
Y
Z
W
∄𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 𝑌 ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑌 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑌 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑌 𝑔𝑖𝑣𝑒𝑛 𝑊 ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑌 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 𝑋 ∧
∄𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 𝑍 ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑍 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑍𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑍 𝑔𝑖𝑣𝑒𝑛 𝑊 ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑍 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 𝑋
Formula ΦCI pattern 𝐶1
CI pattern 𝐶2
Dataset 𝐷2
Dataset 𝐷1
Summary of solutions
Predict that Y and Z are associated even though they are not measured in the same data set.
![Page 119: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/119.jpg)
119
TEST IF IT WORKS IN REAL DATA.
Y X W𝑋, 𝑊|𝑌 Ind
𝑋, 𝑊|∅ Dep
𝑋, 𝑌|∅ Dep
𝑌, 𝑊|∅ Dep
𝑌, 𝑋|𝑊 Dep
𝑌, 𝑊|𝑋 Dep
X W Z
Common variables
𝑋, 𝑊|𝑍 Ind
𝑋, 𝑊|∅ Dep
𝑋, 𝑍|∅ Dep
𝑍, 𝑊|∅ Dep
𝑍, 𝑊|𝑋 Dep
𝑍, 𝑋|𝑊 Dep
X
Y
Z
W
∄𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 𝑌 ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑌 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑌 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑌 𝑔𝑖𝑣𝑒𝑛 𝑊 ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑌 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 𝑋 ∧
∄𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 𝑍 ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑍 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑍𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 ∅ ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑋 𝑡𝑜 𝑍 𝑔𝑖𝑣𝑒𝑛 𝑊 ∧∃𝑚. 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑛𝑔 𝑝𝑎𝑡ℎ 𝑓𝑟𝑜𝑚 𝑍 𝑡𝑜 𝑊 𝑔𝑖𝑣𝑒𝑛 𝑋
Formula ΦCI pattern 𝐶1
CI pattern 𝐶2
Dataset 𝐷2
Dataset 𝐷1
Summary of solutions
Find data sets 𝐷1, 𝐷2 measuring overlapping variables
Look for patterns 𝐶1, 𝐶2.
Predict that Y, Z are associated.
Find a third data set measuring both Y, Z and test if they are associated.
![Page 120: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/120.jpg)
120
TEST IF IT WORKS ON REAL DATA (SIMULATE SCENARIO)
1.Original Dataset
3.Find X, Y, W in D1 and X, Z, W, in D2 that satisfy 𝐶1, 𝐶2.
Test Y, Z for association
2.Split to D1 , D2 and Dtest
containing different samples
Predict Y, Z are associated
Restrict inferences only to cases where the probability of errors is small, i.e. p-values are extreme.
D2
D1
Dtest
D2 Y
ZX W
X W
D1
Dtest Y ZX W
pXY.Z < 0.05 accept 𝐷𝑒𝑝 X, Y 𝐙pXY.Z > 0.3 accept 𝐼𝑛𝑑 X, Y 𝐙Else, undecided (forgo making any inferences)
[Tsamardinos, Triantafillou and Lagani, JMLR 2012]
![Page 121: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/121.jpg)
121
DATASETSName # instances # variables Group Size Variables type Scientific domain
Covtype 581012 55 55 Nominal/Ordinal Agricultural
Read 681 26 26 Nominal/Continuous/Ordinal Business
Infant-mortality 5337 83 83 Nominal Clinical study
Compactiv 8192 22 22 Continuous Computer science
Gisette 7000 5000 50 Continuous Digit recognition
Hiva 4229 1617 50 Nominal Drug discovering
Breast-Cancer 286 17816 50 Continuous Gene expression
Lymphoma 237 7399 50 Continuous Gene expression
Wine 4898 12 12 Continuous Industrial
Insurance-C 9000 84 84 Nominal/Ordinal Insurance
Insurance-N 9000 86 86 Nominal/Ordinal Insurance
p53 16772 5408 50 Continuous Protein activity
Ovarian 216 2190 50 Continuous Proteomics
C&C 1994 128 128 Continuous Social science
ACPJ 15779 28228 50 Continuous Text mining
Bibtex 7395 1995 50 Nominal Text mining
Delicious 16105 1483 50 Nominal Text mining
Dexter 600 11035 50 Nominal Text mining
Nova 1929 12709 50 Nominal Text mining
Ohsumed 5000 14373 50 Nominal Text mining
[Tsamardinos, Triantafillou and Lagani, JMLR 2012]
![Page 122: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/122.jpg)
122
DATASETSName # instances # variables Group Size Variables type Scientific domain
Covtype 581012 55 55 Nominal/Ordinal Agricultural
Read 681 26 26 Nominal/Continuous/Ordinal Business
Infant-mortality 5337 83 83 Nominal Clinical study
Compactiv 8192 22 22 Continuous Computer science
Gisette 7000 5000 50 Continuous Digit recognition
Hiva 4229 1617 50 Nominal Drug discovering
Breast-Cancer 286 17816 50 Continuous Gene expression
Lymphoma 237 7399 50 Continuous Gene expression
Wine 4898 12 12 Continuous Industrial
Insurance-C 9000 84 84 Nominal/Ordinal Insurance
Insurance-N 9000 86 86 Nominal/Ordinal Insurance
p53 16772 5408 50 Continuous Protein activity
Ovarian 216 2190 50 Continuous Proteomics
C&C 1994 128 128 Continuous Social science
ACPJ 15779 28228 50 Continuous Text mining
Bibtex 7395 1995 50 Nominal Text mining
Delicious 16105 1483 50 Nominal Text mining
Dexter 600 11035 50 Nominal Text mining
Nova 1929 12709 50 Nominal Text mining
Ohsumed 5000 14373 50 Nominal Text mining
[Tsamardinos, Triantafillou and Lagani, JMLR 2012]
# predictions
222
0
22
135
423
554
1833
7712
4
1839
226
46647
539165
99241
0
1
856
0
0
0
![Page 123: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/123.jpg)
123
HOW DID WE DO?
• About 700000 predictions in 20 datasets.• Accuracy: The percentage of p-values < 0.05.
• May include false positives and exclude false negatives.
![Page 124: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/124.jpg)
124
HOW DID WE DO?
• About 700000 predictions in 20 datasets.• Accuracy: The percentage of p-values < 0.05.
• May include false positives and exclude false negatives.
98% accuracy vs. 16% of random guessing
![Page 125: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/125.jpg)
125
PREDICT CORRELATION STRENGTH 𝝆𝒀𝒁
X Y Z W
26 possible SMCGs.
How strong is the correlation of Y and Z?
![Page 126: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/126.jpg)
126
PREDICT CORRELATION STRENGTH 𝝆𝒀𝒁
X Y Z W
26 possible SMCGs.
How strong is the correlation of Y and Z?
• Assume multivariate normality and interpret SMCG as path diagram.
• Use the (measured) sample correlations • 𝑟𝑌𝑋, 𝑟𝑌𝑊, 𝑟𝑋𝑊 (𝐷1)• 𝑟𝑍𝑋, 𝑟𝑍𝑊, 𝑟𝑋𝑊 (𝐷2)
• Use rules of path analysis to predict 𝑟𝑌𝑍.
![Page 127: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/127.jpg)
127
PREDICT CORRELATION STRENGTH 𝝆𝒀𝒁
X Y Z W
How strong is the correlation of Y and Z?
• Assume multivariate normality and interpret SMCG as path diagram.
• Use the (measured) sample correlations • 𝑟𝑌𝑋, 𝑟𝑌𝑊, 𝑟𝑋𝑊 (𝐷1)• 𝑟𝑍𝑋, 𝑟𝑍𝑊, 𝑟𝑋𝑊 (𝐷2)
• Use rules of path analysis to predict 𝑟𝑌𝑍.
13 models imply
𝑟𝑌𝑍1 ≈
1
2
𝑟𝑋𝑍
𝑟𝑋𝑌+
𝑟𝑌𝑊
𝑟𝑍𝑊
13 models imply
𝑟𝑌𝑍2 ≈
1
2
𝑟𝑋𝑌
𝑟𝑋𝑍+
𝑟𝑍𝑊
𝑟𝑌𝑊
![Page 128: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/128.jpg)
128
PREDICT CORRELATION STRENGTH 𝝆𝒀𝒁
X Y Z W
How strong is the correlation of Y and Z?
• Assume multivariate normality and interpret SMCG as path diagram.
• Use the (measured) sample correlations • 𝑟𝑌𝑋, 𝑟𝑌𝑊, 𝑟𝑋𝑊 (𝐷1)• 𝑟𝑍𝑋, 𝑟𝑍𝑊, 𝑟𝑋𝑊 (𝐷2)
• Use rules of path analysis to predict 𝑟𝑌𝑍.
13 models imply
𝑟𝑌𝑍1 ≈
1
2
𝑟𝑋𝑍
𝑟𝑋𝑌+
𝑟𝑌𝑊
𝑟𝑍𝑊
13 models imply
𝑟𝑌𝑍2 ≈
1
2
𝑟𝑋𝑌
𝑟𝑋𝑍+
𝑟𝑍𝑊
𝑟𝑌𝑊
Only one of
𝒓𝒀𝒁𝟏
, 𝒓𝒀𝒁𝟐
is < 1
![Page 129: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/129.jpg)
129
PREDICT CORRELATION STRENGTH 𝝆𝒀𝒁
X Y Z W
How strong is the correlation of Y and Z?
• Assume multivariate normality and interpret SMCG as path diagram.
• Use the (measured) sample correlations • 𝑟𝑌𝑋, 𝑟𝑌𝑊, 𝑟𝑋𝑊 (𝐷1)• 𝑟𝑍𝑋, 𝑟𝑍𝑊, 𝑟𝑋𝑊 (𝐷2)
• Use rules of path analysis to predict 𝑟𝑌𝑍.
13 models imply
𝑟𝑌𝑍1 ≈
1
2
𝑟𝑋𝑍
𝑟𝑋𝑌+
𝑟𝑌𝑊
𝑟𝑍𝑊
13 models imply
𝑟𝑌𝑍2 ≈
1
2
𝑟𝑋𝑌
𝑟𝑋𝑍+
𝑟𝑍𝑊
𝑟𝑌𝑊
Only one of
𝒓𝒀𝒁𝟏
, 𝒓𝒀𝒁𝟐
is < 1
You can uniquely identify the skeleton of the graph AND predict the correlation coefficient of Y, Z!
![Page 130: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/130.jpg)
130
HOW DID WE DO?
high
low
den
sity
• Clear trend in predicted vs sample correlations.• Also a systematic bias because the predictions have been selected based on the
independence tests.• Correlation of predicted vs sample correlations is 0.89.• Predictions based on large correlations have reduced bias.
![Page 131: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/131.jpg)
131
HOW DID WE DO?
high
low
den
sity
Predicted vs sample correlations over all data sets, grouped by mean absolute value of the denominators used in their computations
• Clear trend in predicted vs sample correlations• Also a systematic bias because the predictions have been selected based on the
independence tests• Correlation of predicted vs sample correlations is 0.89• Predictions based on large correlations have reduced bias.
![Page 132: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/132.jpg)
-How can you reason with this set of models?You can use do-calculus and estimate (a population of) causal effects.
-Is it useful? Do you make additional inferences than analyzing each data set in isolation?You can make non-trivial inferences, quantitative with additional assumptions.
QUESTIONS
132
![Page 133: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/133.jpg)
1. Integrative causal discoveryi. Motivation.
ii. Causal models.
iii. m-separation.
iv. Reverse engineering causal models (single data set).
v. Problem formulation: Reverse engineering causal models from multiple heterogeneous data sets.
vi. Modeling interventions/selection.
2. Logic-based causal discoveryi. Converting path constraints to logic formulae.
ii. Problem complexity.
iii. Conflict resolution.
iv. Existing algorithms.
v. Reasoning with logic based causal discovery.
vi. Non-trivial inferences-validation.
OUTLINE
133
![Page 134: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/134.jpg)
Integrative logic-based causal discovery.
Different data distributions, same causal mechanism: use causal modeling to connect.
Can handle datasets of different variable sets, different experimental conditions, prior causal knowledge.
Identify the set of causal graphs that simultaneously fit all datasets and reason with this set.
Convert problem to SAT or ASP; exploit 40 years of SAT-solving technology.
Query-based approach to avoid explosion of possible solutions!
Vision of automatically analyzing a large portion of available datasets in a domain.
KEY-POINTS
134
![Page 135: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/135.jpg)
Improving scalability.
Improving quality of learning and robustness.
Further removing restrictive assumptions (e.g., Faithfulness).
Making quantitative predictions.
Extensions for temporal data.
Additional constraints (e.g. Verma constraints).
Feature selection from multiple data sets.
Apply it to real problems.
WHAT IS NEXT?
![Page 136: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/136.jpg)
1. D. Margaritis and F. Bromberg, Efficient Markov Network Discovery Using Particle Filters, Computational Intelligence (2009).
2. S. Triantafillou, I Tsamardinos and IG Tollis, Learning Causal Structure from Overlapping Variable Sets, AISTATS 2010.
3. G. Borboudakis, S. Triantafillou, V. Lagani, I. Tsamardinos, A Constraint-based Approach to Incorporating Prior Knowledge in Causal Models, ESANN 2011.
4. T. Claassen and T. Heskes, A Logical Characterization of Constraint-based Causal Discovery. UAI 2011.
5. T. Claassen and T. Heskes, A Bayesian Approach to Constraint-based Causal Inference, UAI 2012.
6. I. Tsamardinos, S. Triantafillou, V. Lagani, Towards Integrative Causal Analysis of Heterogeneous Data Sets and Studies, Journal of Machine Learning Research (2012).
7. A. Hyttinen, P. O. Hoyer, F. Eberhardt, and M. Järvisalo, Discovering Cyclic Causal Models with Latent Variables: A General SAT-Based Procedure , UAI 2013.
REFERENCES
136
![Page 137: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/137.jpg)
8. A. Hyttinen, F. Eberhardt, and M. Järvisalo, Constraint-based Causal Discovery: Conflict Resolution with Answer Set Programming, UAI 2014.
9. S. Triantafillou, I. Tsamardinos, A. Roumpelaki, Learning Neighborhoods of High Confidence in Constraint-Based Causal Discovery, PGM 2014.
10. S. Triantafillou and I. Tsamardinos, Constraint-based Causal Discovery from Multiple Interventions over Overlapping Variable Sets, Journal of Machine Learning Research (2015).
11. D. Sonntag, M. Järvisalo, Jose M. Pena, A. Hyttinen, Learning Optimal Chain Graphs with Answer Set Programming, UAI 2015.
12. A. Hyttinen, S. Plis, M. Järvisalo, F. Eberhardt, and D. Danks, Causal Discovery from Subsampled Time Series Data by Constraint Optimization, submitted.
13. G. Borboudakis and I. Tsamardinos. Towards Robust and Versatile Causal Discovery for Business Applications. KDD 2016.
14. S. Magliacane, T. Claassen, J.M. Mooij, Ancestral Causal Inference, arXiv:1606.07035
REFERENCES(2)
137
![Page 138: LOGIC-BASED INTEGRATIVE CAUSAL DISCOVERYauai.org/uai2016/tutorials/causal.pdf · 10 SEMI MARKOV CAUSAL GRAPHS directly causes and share a latent common cause Semi Markov Causal Graph](https://reader034.fdocuments.in/reader034/viewer/2022050508/5f9952225b61282e892885d4/html5/thumbnails/138.jpg)
ACKNOWLEDGEMENTS
138
Mens x machina group, University of Crete.Jan Lemeire, Frederick Eberhardt, Antti Hyttinen, Joris MooijERC CAUSALPATH