Download - A Library of Components for Classification Problem Solving

Transcript
Page 1: A Library of Components for  Classification Problem Solving

A Library of Components for Classification Problem Solving

Wenjin Lu and Enrico MottaKnowledge Media Institute

Page 2: A Library of Components for  Classification Problem Solving

Four Main Goals

• To carry out a knowledge-level analysis of classification

• To develop a practical resource to support the development of classification applications

• To provide a concrete set of components to act as a test case for IBROW brokering system and IRS

• To evaluate the UPML framework and the OCML modelling language on a non-trivial test-case

Page 3: A Library of Components for  Classification Problem Solving

UPML Framework

TaskRefiner

Task

PSMRefiner

PSM

Ontol.Refiner

Ontologies

DomainRefiner

DomainModel

PSM-TaskBridge

PSM-DomainBridge

Task-DomainBridge

Page 4: A Library of Components for  Classification Problem Solving

Detailed Modelling in OCML

• Supports domain, task and PSM specification• Large Library (>90 Ontologies)• Extensive experience (~20 projects, 5 years)• Robust Infrastructure

– Both web-based and ‘vanilla’ development environments

• Intg. of specification and operationalization is a good thing! Rapid development and validation Result = both analytical and engineering

resource

Page 5: A Library of Components for  Classification Problem Solving

Amalgamating UPML and OCML

• OCML Base Ontology was revised to comply with UPML Tasks and PSMs become assumption-based

Page 6: A Library of Components for  Classification Problem Solving

Classification

Classification can be seen as the problem of finding the solution (class), which best explains a set of known facts (observables), according to some criterion

Observables

Candidate Sols.

Criterion

Classification Solution

Page 7: A Library of Components for  Classification Problem Solving

Example

Observables

Candidate Sols.

Criterion

Classification Solution

{background=green; area=china...}

Complete-coverage-criterion(every observable has to be explained)

{chinese-granny, dutch-granny, etc..}

{chinese-granny}

Page 8: A Library of Components for  Classification Problem Solving

Observables

Observables = set_of (Observable);Observable = {feature, value}.

Well defined Observables (obs):

({f1, v1} obs {f1, v2} obs) -> v1 = v2

({f1, v1} obs) -> legal_feature_value (f1, v1 )

Page 9: A Library of Components for  Classification Problem Solving

Solutions

Solution = set_of (Feature_Spec);Feature_Spec = {Feature, Feature_value_spec}Feature_value_spec = Unary_Relation

Well defined Solution (sol):{f1, s1} sol holds (s1, v1 ) ->

legal_feature_value (f1, v1 )

Page 10: A Library of Components for  Classification Problem Solving

Matching

Observable={f1, v1} matches Solution=sol iff:

{f1, c} sol holds (c, v1 )

Page 11: A Library of Components for  Classification Problem Solving

Matching Sets of Obs to a Solution

Sol: {{fsol1, c1}...{fsolm, cm}}; Obs: {{fob1, v1}...{fobn, vn}}

Four possible cases: {fj, cj} sol {fj, vj} obs holds (cj, vj)

-> Explained (fj)

{fj, cj} sol {fj, vj} obs not holds (cj, vj) -> Inconsistent(fj)

{fj, vj} obs {fj, cj} sol -> Unexplained (fj)

{fj, vj} obs {fj, cj} sol -> Missing (fj)

Page 12: A Library of Components for  Classification Problem Solving

Default Match Criterion

Match Score:Vector: <I, E, U, M>

Match Comparison RelationS1 = (i1, e1, u1, m1); S2 = (i2, e2, u2, m2)

S1 better_score than S2 iff:

(i1 < i2)

(i2 = i1 e2 < e1) (i2 = i1 e2 = e1 u1 < u2) (i2 = i1 e2 = e1 u2 = u1 m1 < m2)

Page 13: A Library of Components for  Classification Problem Solving

Possible Solution Criteria

• Positive Coverage– Some feature is explained and none is

incosistent

• Complete Coverage– All features are explained and none is

incosistent

Page 14: A Library of Components for  Classification Problem Solving

Hierarchy of Criteria

Solution Criterion

Match Criterion

Match Score Comparison Rel

Macro Score MechanismFeature Score Mechanism

Match Score Mechanism

Page 15: A Library of Components for  Classification Problem Solving

Observables

(def-class observables (set) ?obs "This is simply a set of observables. An important constraint is that there cannot be two values for the same

feature in a set of observables" :iff-def (every ?obs observable) :constraint (not (exists (?ob1 ?ob2) (and (member ?ob1 ?obs) (member ?ob2 ?obs) (has-observable-feature ?ob1 ?f) (has-observable-feature ?ob2 ?f) (has-observable-value ?ob1 ?v1) (has-observable-value ?ob2 ?v2) (not (= ?v1 ?v2))))))

Page 16: A Library of Components for  Classification Problem Solving

Solutions

(def-class solution () ?x "A solution is a set of feature definitions" :iff-def (every ?x feature-definition))

(def-class feature-definition () ?x ((has-feature-name :type feature) (has-feature-value-spec :type unary-relation)) :constraint (=> (and (has-feature-name ?x ?f) (has-feature-value-spec ?x ?spec)) (=> (holds ?spec ?v) (legal-feature-value ?f ?v))))

Page 17: A Library of Components for  Classification Problem Solving

Solution Criterion

(def-class solution-admissibility-criterion () ?c ((applies-to-match-score-type :type match-score-type) (has-solution-admissibility-relation :type unary-relation)) :constraint (=> (and (solution-admissibility-criterion ?c) (has-solution-admissibility-relation ?c ?r) (domain ?r ?d)) (subclass-of ?d match-score)))

Page 18: A Library of Components for  Classification Problem Solving

Monotonicity of Admissibile Solutions

(def-axiom admissibility-is-monotonic "This axiom states that the admissibility criterion is monotonic. That is, if a

solution, ?sol, is admissible, then any solution which is better than ?sol will also be admissible"

(forall (?sol1 ?sol2 ?obs ?criterion) (=> (and (admissible-solution ?sol1 (apply-match-criterion

?criterion ?obs ?sol1) ?criterion)

(better-match-than ?sol2 ?sol1 ?obs ?criterion)) (admissible-solution ?sol2 (apply-match-criterion

?criterion ?obs ?sol2) ?criterion))))

Page 19: A Library of Components for  Classification Problem Solving

Complete Coverage

(def-instance complete-coverage-admissibility-criterion solution-admissibility-criterion ((applies-to-match-score-type default-match-score) (has-solution-admissibility-relation complete-coverage-admissibility-relation)))

(def-relation complete-coverage-admissibility-relation (?score) "a solution should be consistent and explain all features" :constraint (default-match-score ?score) :iff-def (and (= (length (first ?score)) 0) ;;no inconsistency (= (length (third ?score)) 0))) ;;no unexplained

Page 20: A Library of Components for  Classification Problem Solving

Classification Task Ontology

• 42 Definitions• Provides both a theory of classification and a

vocabulary to describe classification problems• Ontology is separated from task specifications

Page 21: A Library of Components for  Classification Problem Solving

Generic Classification Task

• Input roles– Candidate Solutions, Match Criterion, Solution

Criterion, Observables

• Precondition– Both observables and candidate solutions have

to be provided

• Goal– To find a solution from the candidate solutions

which is admissible with respect to the given observables, solution criterion and match criterion

Page 22: A Library of Components for  Classification Problem Solving

Specific Classification Tasks

• Single-Solution Classification Task– Single-solution assumption

• Optimal Classification Tasks– Goal requires optimality

Page 23: A Library of Components for  Classification Problem Solving

Problem Solving Library

• Based on heuristic classification model• Supports both data-directed and solution-

directed classification• Based on search paradigm• Supported by a method ontology

Page 24: A Library of Components for  Classification Problem Solving

Method Ontology: Main Concepts

• Abstractors– Mechanism for performing abstraction on

observables– Abstractor: Obs* -> Obs

• Refiners– Mechanism for specialising a solution– Refiner: Sol -> Sol*

• Candidate Exclusion Criterion– A criterion which is used to decide when a

search path is a dead-end– Default criterion rules out inconsistent solutions

Page 25: A Library of Components for  Classification Problem Solving

Monotonicity of Exclusion Criterion

(def-axiom exclusion-is-monotonic (forall (?sol1 ?sol2 ?obs ?criterion) (=> (and (ruled-out-solution ?sol1 (the-match-score ?sol1) ?criterion) (not (better-match-than ?sol2 ?sol1 ?obs ?criterion))) (ruled-out-solution ?sol2 (the-match-score ?sol2)?criterion))))

Page 26: A Library of Components for  Classification Problem Solving

Axiom of Congruence(def-axiom CONGRUENT-ADMISSIBILITY-AND-EXCLUSION-CRITERIA (forall (?sol ?task) (=> (member ?sol (the-solution-space ?task)) (not (and (admissible-solution ?sol (the-match-score ?sol) (role-value ?task 'has-solution-admissibility-criterion)) (ruled-out-solution ?sol (the-match-score ?sol)

(role-value ?psm

'has-solution-exclusion-criterion)))))))

Page 27: A Library of Components for  Classification Problem Solving

Three Heuristic Classification PSMs

• Two Data-directed– Admissible Solution Classifier

• Finds one admissible solution according to the given criteria• Uses backtracking hill climbing

– Optimal Classifier• Performs complete search looking for optimal solution• Uses best-first strategy• Uses candidate exclusion criterion to prune search space

• One Solution-directed– Goes down the solution hierarchy, acquiring

observables as needed– Ask for observables with max discrimination power

Page 28: A Library of Components for  Classification Problem Solving

Four Assumptions in Main PSMs

• No cycles in abstraction hierarchy• No cycles in refinement hierarchy• At least one class in the solution space is an

admissible solution• The solution refinement hierarchy is consistent

with the candidate exclusion criterion. That is if sol is ruled out, all refinements of sol can also be ruled out

Page 29: A Library of Components for  Classification Problem Solving

Task-Method Hierarchy

abstraction

heuristic-classification-psm

classification

rank-solutions refinement

basic-heuristic-matchselect-abstractor one-step-abstraction collect-refiners apply-refiners

abstraction-psm refinement-psmrank-solutions-psm

Page 30: A Library of Components for  Classification Problem Solving

Example

• Apple Domain– Originally developed in Amsterdam

• Solutions = Apple Types = {granny, noble, delicious...}

• Hierarchy of Apple Types• Features = {bkg-colour, fg-colour, rusty....}• Pretty trivial really!

Page 31: A Library of Components for  Classification Problem Solving

Classification TaskOntology

Heuristic ClassificationOntology

Apple Heuristic ClassificationApplication

Classification TaskSpecification

Classification-to-Class-RepresentationMapping Ontology

AppleDomain Model

Heuristic ClassificationPSMs

Page 32: A Library of Components for  Classification Problem Solving

Mapping Solutions and Obs to Apples

(def-relation-mapping solution :up ((solution ?x) if (or (= ?x apple) (subclass-of ?x apple))))

(def-relation-mapping observable :up ((observable ?x) if (or (== ?X (?f ?v ?obs)) (== ?x (?f ?v)))))

Page 33: A Library of Components for  Classification Problem Solving

More Relation Mappings

(def-relation-mapping has-observable-feature :up ((has-observable-feature ?x ?f) if (or (== ?X (?f ?v ?obs)) (== ?x (?f ?v)))))

(def-relation-mapping has-observable-value :up ((has-observable-value ?x ?v) if (or (== ?X (?f ?v ?obs)) (== ?x (?f ?v)))))

(def-relation-mapping directly-abstracts-from :up ((directly-abstracts-from ?ob ?obs) if (== ?ob (?f ?v ?obs))))

Page 34: A Library of Components for  Classification Problem Solving

Sample Abstractor

(def-instance sugar-abstractor abstractor ((has-body '(lambda (?obs) (in-environment ((?v . (observables-feature-value ?obs 'sugar))) (cond ((>= ?v 70) (list-of 'sweet-level 'high (list-of (list-of 'sugar ?v)))) ((and (< ?v 70) (> ?v 40)) (list-of 'sweet-level 'medium (list-of (list-of 'sugar ?v)))) ((<= ?v 40) (list-of 'sweet-level 'low (list-of (list-of 'sugar ?v)))))))) (applicability-condition (kappa (?obs) (member 'sugar (all-features-in-observables ?obs))))))

Page 35: A Library of Components for  Classification Problem Solving

Generic (reusable) Refiner

(def-instance refinement-through-subclass-of-links refiner "If the solution space is specified by means of classes arranged in a

subclass-of hierarchy, then this is a good refiner to use" ((has-body '(lambda (?sol) (setofall ?sub (direct-subclass-of ?sub ?sol)))) (applicability-condition (kappa (?sol) (and (class ?sol) (exists ?sub (direct-subclass-of ?sub ?sol)))))))

Page 36: A Library of Components for  Classification Problem Solving

Evaluation/Results

• All PSMs successfully tested on the apple domain

• Assumptions also successfully tested in the domain

• Library available online in WebOnto

Page 37: A Library of Components for  Classification Problem Solving
Page 38: A Library of Components for  Classification Problem Solving
Page 39: A Library of Components for  Classification Problem Solving
Page 40: A Library of Components for  Classification Problem Solving
Page 41: A Library of Components for  Classification Problem Solving

Next Tasks

• Start work on Internet Reasoning Service• Approach: Ever increasing levels of intelligent

support– Browsing/Navigation/Manual PSM Configuration– Intelligent Assistant

• Semi-automated component selection/configuration

– Intelligent Broker• Multiple libraries/multiple platforms/symbol-level

interoperability

• Application to more complex domains– Scientific Classification,

Selection of Manufacturing Tech.

Page 42: A Library of Components for  Classification Problem Solving

Possible Platforms for IRS

• Specialized WebOnto Configuration• Protégé

– Intg. Protégé with OCML Library • Collaboration with Stanford (i.e., Monica)

– Dedicated Tabs to support PSM selection/reuse

• New Java/Lisp Tool– Java Applets interfaced with library sitting on

Lisp server

Page 43: A Library of Components for  Classification Problem Solving

Classification Library in OCML (at the end of IBROW 1)

• Task spec (TaskSpec1)

• Flat classification PSM (GenPSM1)

• Applied to apple and Rocky-III domains