@SrcMinistry @MariuszGil

Machine Learning for Developers

Data processing

@SrcMinistry

My story

Data classification

Bot detection

Minimize risk of error

Predictions

Click probability

Maximize CTR or eCPM

A lot of data

data + algo = result

Real problem

+ value estimator

+ chance of sell

+ $ optimization

Tens of thousands historical transactions

Tens of data components

Hundreds of data components

HOW?

Machine Learning Theory

A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E

Tom M. Mitchell

Task

Typical ML techniquesClassification Regression Clustering Dimensionality reduction Association learning

oooo

ooo

ooo oo

o oo

oo o o oo

ooo

oo

o

feature 1

feat

ure

2

oooo

ooo

ooo oo

o oo

oo o o oo

ooo

oo

o

feature 1

feat

ure

2

oooo

ooo

ooo oo

o oo

oo o o oo

ooo

oo

o

feature 1

feat

ure

2

Experience

Typical ML paradigmsSupervised learning Unsupervised learning Reinforcement learning

Accuracy

Substantive Expertise

Hacking skillsMath & Statistics

Knowledge

TraditionalResearch

DangerZone!

MachineLearning

DataScience

Substantive

Expertise

Hacking skillsMath & Statistics

Knowledge

Evil

Outside CommitteeMember

Not that dangerous, in retrospect

MachineLearning

James BondVillain

NSA

Data Science

That Guy Who Stole Your Online Identity Thesis Advisor

Grad School Mate

Machine Learning Practice

data + algo = result

+-------+--------+------+--------+---------+-------+ | brand | model | year | milage | service | price | +-------+--------+------+--------+---------+-------+ | ford | mondeo | 2005 | 123000 | 9900 | 67000 | +-------+--------+------+--------+---------+-------+ | ford | mondeo | 2005 | 175000 | 9900 | 30000 | +-------+--------+------+--------+---------+-------+ | ford | focus | 2010 | 45000 | 6700 | 30000 | +-------+--------+------+--------+---------+-------+

…

Learning Data

Algorithm Learning

Classifier ModelReal Data Classification

Failure recipe

+-------+--------+------+--------+---------+-------+ | brand | model | year | milage | service | price | +-------+--------+------+--------+---------+-------+ | ford | mondeo | 2005 | 123000 | 9900 | 67000 | +-------+--------+------+--------+---------+-------+ | ford | mondeo | 2005 | 175000 | 9900 | 30000 | +-------+--------+------+--------+---------+-------+ | ford | focus | 2010 | 45000 | 6700 | 30000 | +-------+--------+------+--------+---------+-------+

…

+-------+--------+------+--------+---------+--------+-------+ | brand | model | year | milage | service | repair | price | +-------+--------+------+--------+---------+--------+-------+ | ford | mondeo | 2005 | 123000 | 9000 | 900 | 67000 | +-------+--------+------+--------+---------+--------+-------+ | ford | mondeo | 2005 | 175000 | 900 | 9000 | 30000 | +-------+--------+------+--------+---------+--------+-------+ | ford | focus | 2010 | 45000 | 3700 | 3000 | 30000 | +-------+--------+------+--------+---------+--------+-------+

…

+-------+--------+------+--------+---------+--------+-------+ | brand | model | year | milage | service | repair | price | +-------+--------+------+--------+---------+--------+-------+ | ford | mondeo | 2005 | 123000 | 9000 | 900 | 67000 | +-------+--------+------+--------+---------+--------+-------+ | ford | mondeo | 2005 | 175000 | 900 | 9000 | 30000 | +-------+--------+------+--------+---------+--------+-------+ | ford | mondeo | 2005 | 175000 | 900 | 9000 | 45000 | +-------+--------+------+--------+---------+--------+-------+ | ford | focus | 2010 | 45000 | 3700 | 3000 | 30000 | +-------+--------+------+--------+---------+--------+-------+

…

+-------+--------+-----+------+--------+---------+--------+-------+ | brand | model | gen | year | milage | service | repair | price | +-------+--------+-----+------+--------+---------+--------+-------+ | ford | mondeo | 4 | 2005 | 123000 | 9000 | 900 | 67000 | +-------+--------+-----+------+--------+---------+--------+-------+ | ford | mondeo | 3 | 2005 | 175000 | 900 | 9000 | 30000 | +-------+--------+-----+------+--------+---------+--------+-------+ | ford | mondeo | 4 | 2005 | 175000 | 900 | 9000 | 45000 | +-------+--------+-----+------+--------+---------+--------+-------+ | ford | focus | 4 | 2010 | 45000 | 3700 | 3000 | 30000 | +-------+--------+-----+------+--------+---------+--------+-------+

…

+-------+--------+-----+------+--------+---------+--------+------+---------------+-------+ | brand | model | gen | year | milage | service | repair | igla | crying German | price | +-------+--------+-----+------+--------+---------+--------+------+---------------+-------+ | ford | mondeo | 4 | 2005 | 123000 | 9000 | 900 | 0 | 0 | 67000 | +-------+--------+-----+------+--------+---------+--------+------+---------------+-------+ | ford | mondeo | 3 | 2005 | 175000 | 900 | 9000 | 1 | 1 | 30000 | +-------+--------+-----+------+--------+---------+--------+------+---------------+-------+ | ford | mondeo | 4 | 2005 | 175000 | 900 | 9000 | 0 | 0 | 45000 | +-------+--------+-----+------+--------+---------+--------+------+---------------+-------+ | ford | focus | 4 | 2010 | 45000 | 3700 | 3000 | 1 | 0 | 30000 | +-------+--------+-----+------+--------+---------+--------+------+---------------+-------+

…

Understand your data first

Exploratory analysis

http://blogs.adobe.com/digitalmarketing/wp-content/uploads/2013/08/aq2.jpg

ML pipeline

Raw Data Collection

Pre-processing

Sampling

Training Dataset

Algorithm Training

Optimization

Post-processing

Final model

Pre-processingFeature Selection

Feature Scaling

Dimensionality Reduction

Performance Metrics

Model Selection

Test Dataset

Cro

ss V

alid

atio

n

Final ModelEvaluation

Pre-processing Classification

Missing Data

Feature Extraction

DataSplit

Data

Raw Data Collection

Pre-processing

Sampling

Training Dataset

Algorithm Training

Optimization

Final model

Pre-processingFeature Selection

Feature Scaling

Dimensionality Reduction

Performance Metrics

Model Selection

Test Dataset

Cro

ss V

alid

atio

n

Final ModelEvaluation

Pre-processing Classification

Missing Data

Feature Extraction

DataSplit

Post-processing

Data

Classification algorithmsLinear Classification Logistic Regression Linear Discriminant Analysis PLS Discriminant Analysis

Non-Linear Classification Mixture Discriminant Analysis Quadratic Discriminant Analysis Regularized Discriminant Analysis Neural Networks Flexible Discriminant Analysis Support Vector Machines k-Nearest Neighbor Naive Bayes

Decission Trees for Classification Classification and Regression Trees C4.5 PART Bagging CART Random Forest Gradient Booster Machines Boosted 5.0

Regression algorithmsLinear Regiression Ordinary Least Squares Regression Stepwise Linear Regression Prinicpal Component Regression Partial Least Squares Regression

Non-Linear Regression / Penalized Regression Ridge Regression Least Absolute Shrinkage ElasticNet Multivariate Adaptive Regression Support Vector Machines k-Nearest Neighbor Neural Network

Decission Trees for Regression Classification and Regression Trees Conditional Decision Tree Rule System Bagging CART Random Forest Gradient Boosted Machine Cubist

Algorithm is only element in the ML chain

Demo #1

> dataset(iris)

> head(iris)

Sepal.Length Sepal.Width Petal.Length Petal.Width Species 1 5.1 3.5 1.4 0.2 setosa 2 4.9 3.0 1.4 0.2 setosa 3 4.7 3.2 1.3 0.2 setosa 4 4.6 3.1 1.5 0.2 setosa 5 5.0 3.6 1.4 0.2 setosa 6 5.4 3.9 1.7 0.4 setosa

> tail(iris) Sepal.Length Sepal.Width Petal.Length Petal.Width Species 145 6.7 3.3 5.7 2.5 virginica 146 6.7 3.0 5.2 2.3 virginica 147 6.3 2.5 5.0 1.9 virginica 148 6.5 3.0 5.2 2.0 virginica 149 6.2 3.4 5.4 2.3 virginica 150 5.9 3.0 5.1 1.8 virginica

> plot(iris[,1:4])

> library(mclust)

> class = iris$Species

> mod2 = MclustDA(iris[,1:4], class, modelType = „EDDA")

> table(class)

class setosa versicolor virginica 50 50 50

> summary(mod2)

------------------------------------------------ Gaussian finite mixture model for classification ------------------------------------------------

EDDA model summary:

log.likelihood n df BIC -187.7097 150 36 -555.8024 Classes n Model G setosa 50 VEV 1 versicolor 50 VEV 1 virginica 50 VEV 1

Training classification summary:

Predicted Class setosa versicolor virginica setosa 50 0 0 versicolor 0 47 3 virginica 0 0 50

Training error = 0.02

> plot(mod2, what = "scatterplot")

Demo #2

> head(titanic.raw)

Class Sex Age Survived 1 3rd Male Child No 2 3rd Male Child No 3 3rd Male Child No 4 3rd Male Child No 5 3rd Male Child No 6 3rd Male Child No

> tail(titanic.raw)

Class Sex Age Survived 2196 Crew Female Adult Yes 2197 Crew Female Adult Yes 2198 Crew Female Adult Yes 2199 Crew Female Adult Yes 2200 Crew Female Adult Yes 2201 Crew Female Adult Yes

> summary(titanic.raw)

Class Sex Age Survived 1st :325 Female: 470 Adult:2092 No :1490 2nd :285 Male :1731 Child: 109 Yes: 711 3rd :706 Crew:885

> library(arules)

Ładowanie wymaganego pakietu: Matrix Dołączanie pakietu: ‘arules’

> rules <- apriori(titanic.raw)

Apriori

Parameter specification: confidence minval smax arem aval originalSupport support minlen maxlen target ext 0.8 0.1 1 none FALSE TRUE 0.1 1 10 rules FALSE

Algorithmic control: filter tree heap memopt load sort verbose 0.1 TRUE TRUE FALSE TRUE 2 TRUE

Absolute minimum support count: 220

set item appearances ...[0 item(s)] done [0.00s]. set transactions ...[10 item(s), 2201 transaction(s)] done [0.00s]. sorting and recoding items ... [9 item(s)] done [0.00s]. creating transaction tree ... done [0.00s]. checking subsets of size 1 2 3 4 done [0.00s]. writing ... [27 rule(s)] done [0.00s]. creating S4 object ... done [0.00s].

> rules <- apriori(titanic.raw, + parameter = list(minlen=2, supp=0.005, conf=0.8), + appearance = list(rhs=c("Survived=No", "Survived=Yes"), + default="lhs"), + control = list(verbose=F))

> rules.sorted <- sort(rules, by="lift")

> subset.matrix <- is.subset(rules.sorted, rules.sorted)

> subset.matrix[lower.tri(subset.matrix, diag=T)] <- NA

> redundant <- colSums(subset.matrix, na.rm=T) >= 1

> which(redundant)

{Class=2nd,Sex=Female,Age=Child,Survived=Yes} 2 {Class=1st,Sex=Female,Age=Adult,Survived=Yes} 4 {Class=Crew,Sex=Female,Age=Adult,Survived=Yes} 7 {Class=2nd,Sex=Female,Age=Adult,Survived=Yes} 8

> rules.pruned <- rules.sorted[!redundant]

> inspect(rules.pruned)

lhs rhs support confidence lift 1 {Class=2nd,Age=Child} => {Survived=Yes} 0.010904134 1.0000000 3.095640 4 {Class=1st,Sex=Female} => {Survived=Yes} 0.064061790 0.9724138 3.010243 2 {Class=2nd,Sex=Female} => {Survived=Yes} 0.042253521 0.8773585 2.715986 5 {Class=Crew,Sex=Female} => {Survived=Yes} 0.009086779 0.8695652 2.691861 9 {Class=2nd,Sex=Male,Age=Adult} => {Survived=No} 0.069968196 0.9166667 1.354083 3 {Class=2nd,Sex=Male} => {Survived=No} 0.069968196 0.8603352 1.270871 12 {Class=3rd,Sex=Male,Age=Adult} => {Survived=No} 0.175829169 0.8376623 1.237379 6 {Class=3rd,Sex=Male} => {Survived=No} 0.191731031 0.8274510 1.222295

Applications

Tools

Benefits & Problems

oooo

ooo

ooo oo

o oo

oo o o oo

ooo

oo

o

feature 1

feat

ure

2

o

o

Does it do well onthe training data?

Does it do well onthe test data?

Better features /Better parameters

More data

Done!

No No

Yes

by Andrew Ng

Understand your needs first

Tools will change Ideas are immortal

@SrcMinistry

Thanks!

@MariuszGil