Intégration de connaissances ontologiques pour l apprentissage des réseaux bayésiens

Transcription

1 Bayesian networks Ontologies BNs and Ontologies Intégration de connaissances ontologiques pour l apprentissage des réseaux bayésiens Montassar Ben Messaoud 12, Mouna Ben Ishak 12, Philippe Leray 2, Nahla Ben Amor 1 1 LARODEC, ISG, Université de Tunis, Tunisie 2 Connaissances et Décision, LINA UMR 6241, Nantes, France Colloque Apprentissage Artificiel & Fouille de Données 28 juin 2012, Univ. Paris 13 Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 1/22

2 Bayesian networks Ontologies BNs and Ontologies Motivations Knowledge-based systems aim to make expertise available for decision making and information sharing To resolve some complex problems, the combination of different knowledge-based systems can be very powerful Our idea Combine Two KBSs : Bayesian Networks and Ontologies, to help both Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 2/22

3 Bayesian networks Ontologies BNs and Ontologies Motivations Knowledge-based systems aim to make expertise available for decision making and information sharing To resolve some complex problems, the combination of different knowledge-based systems can be very powerful Our idea Combine Two KBSs : Bayesian Networks and Ontologies, to help both Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 2/22

4 Bayesian networks Ontologies BNs and Ontologies Motivations Knowledge-based systems aim to make expertise available for decision making and information sharing To resolve some complex problems, the combination of different knowledge-based systems can be very powerful Our idea Combine Two KBSs : Bayesian Networks and Ontologies, to help both Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 2/22

5 Bayesian networks Ontologies BNs and Ontologies Outline... 1 Bayesian networks BN definition BN learning 2 Ontologies Ontology definition Ontology learning 3 BNs and Ontologies Existing work Our proposal Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 3/22

6 Bayesian networks Ontologies BNs and Ontologies Bayesian network definition Definition [Pearl, 1985] A Bayesian network (BN) is defined by one qualitative description of (conditional) dependences / independences between variables directed acyclic graph (DAG) one quantitative description of these dependences conditional probability distributions (CPDs) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 4/22

7 Bayesian networks Ontologies BNs and Ontologies Example one topological order : B, E, A, R, T (not unique) P(Burglary)=[ ] P(Earthquake)=[ ] B Burglary E Earthquake P(Radio Earthquake) Earthquake = Y N Radio=Y Radio=N P(TV Radio) A Alarm P(Alarm Burglary,Earthquake) R Radio Radio = Y N TV=Y TV=N Burglary,Earthquake = Y,Y Y,N N,Y N,N Alarm=Y Alarm=N Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 5/22 T TV

8 Bayesian networks Ontologies BNs and Ontologies Consequence Chain rule P(S) = P(S 1 ) P(S 2 S 1 ) P(S 3 S 1, S 2 ) P(S n S 1... S n 1 ) Consequence with a BN P(S i S 1... S i 1 ) = P(S i parents(s i )) so P(S) = Π n i=1 P(S i parents(s i )) The (global) joint probability distribution is decomposed in a product of (local) conditional distributions BN = compact representation of the joint distribution P(S) given some information about dependence relationships between variables Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 6/22

9 Bayesian networks Ontologies BNs and Ontologies Consequence Chain rule P(S) = P(S 1 ) P(S 2 S 1 ) P(S 3 S 1, S 2 ) P(S n S 1... S n 1 ) Consequence with a BN P(S i S 1... S i 1 ) = P(S i parents(s i )) so P(S) = Π n i=1 P(S i parents(s i )) The (global) joint probability distribution is decomposed in a product of (local) conditional distributions BN = compact representation of the joint distribution P(S) given some information about dependence relationships between variables Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 6/22

10 Bayesian networks Ontologies BNs and Ontologies Markov equivalence Definition B 1 and B 2 are Markov equivalent iff both describe exactly the same conditional (in)dependence statements Graphical properties B 1 and B 2 have the same skeleton, V-structures and inferred edges All the equivalent graphs (= equivalence class) can be summarized by one partially directed DAG named CPDAG or Essential Graph Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 7/22

11 Bayesian networks Ontologies BNs and Ontologies Markov equivalence Definition B 1 and B 2 are Markov equivalent iff both describe exactly the same conditional (in)dependence statements Graphical properties B 1 and B 2 have the same skeleton, V-structures and inferred edges All the equivalent graphs (= equivalence class) can be summarized by one partially directed DAG named CPDAG or Essential Graph Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 7/22

12 Bayesian networks Ontologies BNs and Ontologies Markov equivalence A S A S T L B T L B O O X D X D Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 8/22

13 Bayesian networks Ontologies BNs and Ontologies BN structure learning How to build / learn a Bayesian Network? 1 DAG is known, how to determine the CPDs? from experts : knowledge elicitation from complete data / incomplete data 2 DAG is unknown, how to determine it (or the CPDAG)? from complete data / incomplete data Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 9/22

14 Bayesian networks Ontologies BNs and Ontologies Structure learning is a complex task Size of the solution space The number of possible DAGs with n variables is super-exponential w.r.t n (Robinson 77) { NS(n) = 1, n = 0 or 1 n i=1 ( 1)i+1( n i ) 2 i(n 1) NS(n i), n > 1 NS(5) = NS(10) = An exhaustive search is impossible! One thousand millenniums = seconds Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 10/22

15 Bayesian networks Ontologies BNs and Ontologies Structure learning algorithms How to search a good BN? Constraint-based methods BN = independence model find CI in data in order to build the DAG Score-based methods BN = probabilistic model that must fit data as well as possible search the DAG space in order to maximize a scoring function Hybrid methods Some issues Identifiability : learning algorithms can t distinguish between any structure in the same equivalence class more Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 11/22

16 Bayesian networks Ontologies BNs and Ontologies Structure learning algorithms How to search a good BN? Constraint-based methods BN = independence model find CI in data in order to build the DAG Score-based methods BN = probabilistic model that must fit data as well as possible search the DAG space in order to maximize a scoring function Hybrid methods Some issues Learning algorithms can deal with a priori knowledge in order to reduce search space : white list, black list,node ordering, repetition of local structures... more Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 11/22

17 Bayesian networks Ontologies BNs and Ontologies Outline... 1 Bayesian networks BN definition BN learning 2 Ontologies Ontology definition Ontology learning 3 BNs and Ontologies Existing work Our proposal Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 12/22

18 Bayesian networks Ontologies BNs and Ontologies Ontology Basics Definition [Gruber, 1993] shared understanding within a community of people declarative specification of entities and their relationships with each other separate the domain knowledge from the operational knowledge Construction/Evolution: expertise and/or machine learning Reasoning: description logic reasoners Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 13/22

19 Bayesian networks Ontologies BNs and Ontologies Elements of an ontology C: Classes (concepts) P: Attributes (properties) H: Hierarchical structure (is-a, part-of relations) R: Other semantic relationships I: Instances (individuals) Fire Flood Landslide Catastrophes Tsunami Volcano A: Axioms (logic statements) Earthquake Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 14/22

20 Bayesian networks Ontologies BNs and Ontologies Elements of an ontology C: Classes (concepts) Landslide P: Attributes (properties) H: Hierarchical structure (is-a, part-of relations) Fire Catastrophes Tsunami R: Other semantic relationships I: Instances (individuals) Flood Name Localization Date Nb_killed Volcano A: Axioms (logic statements) Earthquake Name Mgnitude Localization Date Nb_killed Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 14/22

21 Bayesian networks Ontologies BNs and Ontologies Elements of an ontology C: Classes (concepts) P: Attributes (properties) Catastrophes H: Hierarchical structure (is-a, part-of relations) R: Other semantic relationships is-a Man-made is-a Natural I: Instances (individuals) is-a is-a is-a is-a is-a is-a A: Axioms (logic statements) Flood Fire Earthquake Landslide Volcano Tsunami Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 14/22

22 Bayesian networks Ontologies BNs and Ontologies Elements of an ontology C: Classes (concepts) P: Attributes (properties) H: Hierarchical structure (is-a, part-of relations) R: Other semantic relationships I: Instances (individuals) A: Axioms (logic statements) Catastrophes is-a is-a Man-made Natural is-a is-a is-a is-a is-a is-a Flood Fire Earthquake Landslide Volcano Tsunami Causes Causes Causes Causes Causes Causes Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 14/22

23 Bayesian networks Ontologies BNs and Ontologies Elements of an ontology C: Classes (concepts) Catastrophes P: Attributes (properties) H: Hierarchical structure (is-a, part-of relations) is-a Man-made is-a Natural R: Other semantic relationships is-a is-a is-a is-a is-a is-a I: Instances (individuals) Flood Fire Earthquake Landslide Volcano Tsunami A: Axioms (logic statements) is-instance is-instance is-instance Aleppo Shaanxi Haiti Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 14/22

24 Bayesian networks Ontologies BNs and Ontologies Elements of an ontology C: Classes (concepts) Catastrophes P: Attributes (properties) H: Hierarchical structure (is-a, part-of relations) is-a Man-made is-a Natural R: Other semantic relationships is-a is-a is-a is-a is-a is-a I: Instances (individuals) Flood Fire Earthquake Landslide Volcano Tsunami A: Axioms (logic statements) is-instance is-instance is-instance Aleppo Shaanxi Haiti Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 14/22

25 Bayesian networks Ontologies BNs and Ontologies Ontology learning : subtasks Ontology population Get new instances of concept(s) already present in the ontology Ontology enrichment Update (add or modify) concepts, properties and relations in a given ontology Evolution vs. revolution Evolution (ontology continuity): Add new knowledge Revolution (ontology discontinuity): Modify existing knowledge Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 15/22

26 Bayesian networks Ontologies BNs and Ontologies Ontology learning : subtasks Ontology population Get new instances of concept(s) already present in the ontology Ontology enrichment Update (add or modify) concepts, properties and relations in a given ontology Evolution vs. revolution Evolution (ontology continuity): Add new knowledge Revolution (ontology discontinuity): Modify existing knowledge Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 15/22

27 Bayesian networks Ontologies BNs and Ontologies Ontology learning : subtasks Ontology population Get new instances of concept(s) already present in the ontology Ontology enrichment Update (add or modify) concepts, properties and relations in a given ontology Evolution vs. revolution Evolution (ontology continuity): Add new knowledge Revolution (ontology discontinuity): Modify existing knowledge Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 15/22

28 Bayesian networks Ontologies BNs and Ontologies Ontology learning : methods [Buitelaar & Cimiano, 2008] Bridging the gap between text and knowledge Natural Language Processing [Buitelaar et al., 2003, Velardi et al., 2005] Concept extraction Taxonomy learning (is-a, part-of) Population (information extraction) Machine Learning Clustering for taxonomy learning [Bisson et al., 2000] Association rules for relation discovery [Madche & Staab, 2000] ILP for relation discovery [Rudolph et al., 2007] Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 16/22

29 Bayesian networks Ontologies BNs and Ontologies Ontology learning : methods [Buitelaar & Cimiano, 2008] Bridging the gap between text and knowledge Natural Language Processing [Buitelaar et al., 2003, Velardi et al., 2005] Concept extraction Taxonomy learning (is-a, part-of) Population (information extraction) Machine Learning Clustering for taxonomy learning [Bisson et al., 2000] Association rules for relation discovery [Madche & Staab, 2000] ILP for relation discovery [Rudolph et al., 2007] Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 16/22

30 Bayesian networks Ontologies BNs and Ontologies Ontology learning : methods [Buitelaar & Cimiano, 2008] Bridging the gap between text and knowledge Natural Language Processing [Buitelaar et al., 2003, Velardi et al., 2005] Concept extraction Taxonomy learning (is-a, part-of) Population (information extraction) Machine Learning Clustering for taxonomy learning [Bisson et al., 2000] Association rules for relation discovery [Madche & Staab, 2000] ILP for relation discovery [Rudolph et al., 2007] Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 16/22

31 Bayesian networks Ontologies BNs and Ontologies Outline... 1 Bayesian networks BN definition BN learning 2 Ontologies Ontology definition Ontology learning 3 BNs and Ontologies Existing work Our proposal Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 17/22

32 Bayesian networks Ontologies BNs and Ontologies Existing work Bayesian Network = Ontology BayesOWL [Ding & Peng, 2004] OntoBayes [Yang & Calmet, 2005] PR-OWL [Costa & Laskey, 2006] Use of BNs for probabilistic modeling and reasoning (no learning) Ontology = Bayesian Network BN basic construction using ontologies [Devitt et al., 2006] Ontology-based semi-automatic construction of BN in E-health applications [Jeon & Ko, 2007] Use of ontologies to manually build a BN (without data) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 18/22

33 Bayesian networks Ontologies BNs and Ontologies Existing work Bayesian Network = Ontology BayesOWL [Ding & Peng, 2004] OntoBayes [Yang & Calmet, 2005] PR-OWL [Costa & Laskey, 2006] Use of BNs for probabilistic modeling and reasoning (no learning) Ontology = Bayesian Network BN basic construction using ontologies [Devitt et al., 2006] Ontology-based semi-automatic construction of BN in E-health applications [Jeon & Ko, 2007] Use of ontologies to manually build a BN (without data) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 18/22

34 Bayesian networks Ontologies BNs and Ontologies Existing work Bayesian Network = Ontology BayesOWL [Ding & Peng, 2004] OntoBayes [Yang & Calmet, 2005] PR-OWL [Costa & Laskey, 2006] Use of BNs for probabilistic modeling and reasoning (no learning) Ontology = Bayesian Network BN basic construction using ontologies [Devitt et al., 2006] Ontology-based semi-automatic construction of BN in E-health applications [Jeon & Ko, 2007] Use of ontologies to manually build a BN (without data) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 18/22

35 Bayesian networks Ontologies BNs and Ontologies Existing work Bayesian Network = Ontology BayesOWL [Ding & Peng, 2004] OntoBayes [Yang & Calmet, 2005] PR-OWL [Costa & Laskey, 2006] Use of BNs for probabilistic modeling and reasoning (no learning) Ontology = Bayesian Network BN basic construction using ontologies [Devitt et al., 2006] Ontology-based semi-automatic construction of BN in E-health applications [Jeon & Ko, 2007] Use of ontologies to manually build a BN (without data) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 18/22

36 Bayesian networks Ontologies BNs and Ontologies Our proposal Bridging the gap between text and knowledge Ontology + data = Bayesian Network = Ontology BN Structure Learning 2 variations Ontology Evolution 1 Causal BNs more semantical causal discovery SemCaDo 1.0 [Ben Messaoud & al. 2009] causal discovery for ontology evolution SemCaDo 2.0 [Ben Messaoud & al., 2009] more 2 Object-Oriented BNs OOBN structure learning and ontology evolution O2C [Ben Ishak et al., 2011] more Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 19/22

37 Bayesian networks Ontologies BNs and Ontologies Our proposal Bridging the gap between data and knowledge Ontology + data = Bayesian Network = Ontology BN Structure Learning 2 variations Ontology Evolution 1 Causal BNs more semantical causal discovery SemCaDo 1.0 [Ben Messaoud & al. 2009] causal discovery for ontology evolution SemCaDo 2.0 [Ben Messaoud & al., 2009] more 2 Object-Oriented BNs OOBN structure learning and ontology evolution O2C [Ben Ishak et al., 2011] more Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 19/22

38 Bayesian networks Ontologies BNs and Ontologies Our proposal Bridging the gap between data and knowledge Ontology + data = Bayesian Network = Ontology BN Structure Learning 2 variations Ontology Evolution 1 Causal BNs more semantical causal discovery SemCaDo 1.0 [Ben Messaoud & al. 2009] causal discovery for ontology evolution SemCaDo 2.0 [Ben Messaoud & al., 2009] more 2 Object-Oriented BNs OOBN structure learning and ontology evolution O2C [Ben Ishak et al., 2011] more Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 19/22

39 Bayesian networks Ontologies BNs and Ontologies Our proposal Bridging the gap between data and knowledge Ontology + data = Bayesian Network = Ontology BN Structure Learning 2 variations Ontology Evolution 1 Causal BNs more semantical causal discovery SemCaDo 1.0 [Ben Messaoud & al. 2009] causal discovery for ontology evolution SemCaDo 2.0 [Ben Messaoud & al., 2009] more 2 Object-Oriented BNs OOBN structure learning and ontology evolution O2C [Ben Ishak et al., 2011] more Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 19/22

40 Bayesian networks Ontologies BNs and Ontologies Conclusion BN structure learning is NP hard, need of prior knowledge Ontology evolution is difficult, based on text-mining Cooperation can help for both tasks Originality of our proposal BN structure learning : use ontology instead of expert knowledge : separation between expert acquisition and structure learning Ontology evolution : use BN structure learning to directly discover relationships from data Difficulties No similar work or benchmark for a comparative study SemCaDo : one concept = attribute = node O2C : OOBN learning is too complex Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 20/22

41 Bayesian networks Ontologies BNs and Ontologies Conclusion BN structure learning is NP hard, need of prior knowledge Ontology evolution is difficult, based on text-mining Cooperation can help for both tasks Originality of our proposal BN structure learning : use ontology instead of expert knowledge : separation between expert acquisition and structure learning Ontology evolution : use BN structure learning to directly discover relationships from data Difficulties No similar work or benchmark for a comparative study SemCaDo : one concept = attribute = node O2C : OOBN learning is too complex Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 20/22

42 Bayesian networks Ontologies BNs and Ontologies Conclusion BN structure learning is NP hard, need of prior knowledge Ontology evolution is difficult, based on text-mining Cooperation can help for both tasks Originality of our proposal BN structure learning : use ontology instead of expert knowledge : separation between expert acquisition and structure learning Ontology evolution : use BN structure learning to directly discover relationships from data Difficulties No similar work or benchmark for a comparative study SemCaDo : one concept = attribute = node O2C : OOBN learning is too complex Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 20/22

43 Bayesian networks Ontologies BNs and Ontologies Future works What next? Creation of benchmarks Generalize SemCaDo or O2C with more general models Ideas : Multi Entity BNs [Laskey, 2006] Relational BNs [Getoor et al., 2007] Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 21/22

44 Bayesian networks Ontologies BNs and Ontologies Thank you for your attention Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 22/22

45 Outline... 4 Bayesian network Structure Learning constraint-based methods score-based methods local search methods 5 Causal Bayesian Networks definition causal BN structure learning 6 SemCaDo algorithm definition experimental study 7 O2C algorithm definition algorithm Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 1/49

46 Constraint-based methods How to search a good BN? Constraint-based methods BN = independence model find CI in data in order to build the DAG Score-based methods BN = probabilistic model that must fit data as well as possible search the DAG space in order to maximize a scoring function Hybrid methods Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 2/49

47 Constraint-based methods Two reference algorithms Pearl et Verma : IC, IC* Spirtes, Glymour et Scheines : SGS, PC, CI, FCI Common principle Build an undirected graph describing direct dependences between variables (χ 2 tests) by adding edges (Pearl et Verma) by deleting edges (SGS) Detect V-structures (from previous statistical tests) Propagate some edge orientation (inferred edges) in order to obtain a CPDAG Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 3/49

48 constraint-based methods Some inconvenients Reliability of CI test conditionally to several variables with a limited amount of data) SGS heuristic : if df < N 10, then declare dependence Combinatorial explosion of the number of tests PC heuristic : begin with order 0 (X A X B ) then order 1 (X A X B X C ), etc... Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 4/49

49 PC algorithm Step 0 : undirected complete graph Left : target BN used to generate 5000 samples A S A S T L B T L B O O X D X D Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 5/49

50 PC algorithm Step 1a : delete all order 0 independences discovered χ 2 : S A L A B A O A X A D A T S L T O B X B A S A S T L B T L B O O X D X D Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 5/49

51 PC algorithm Step 1a : delete all order 1 independences discovered χ 2 : T A O O S L X S L B T S X T O D T O... A S A S T L B T L B O O X D X D Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 5/49

52 PC algorithm Step 1a : delete all order 2 independences discovered χ 2 : D S {L, B} X O {T, L} D O {T, L} A S A S T L B T L B O O X D X D Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 5/49

53 PC algorithm Step 2 : research V-structures χ 2 : one V-structure T O L is discovered A S A S T L B T L B O O X D X D Step 3 : inferred edges no one in this example Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 5/49

54 PC algorithm From CPDAG to DAG Orientation of the remaining undirected edges (only constraint : do not create any new V-structure) A S A S T L B T L B O O X D X D Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 5/49

55 PC algorithm Obtained DAG versus target one χ 2 test with 5000 samples fails to discover A T, O X and O D A S A S T L B T L B O O X D X D Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 5/49

56 Score-based methods How to search a good BN? Constraint-based methods BN = independence model find CI in data in order to build the DAG Score-based methods BN = probabilistic model that must fit data as well as possible search the DAG space in order to maximize a scoring function Hybrid methods Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 6/49

57 Notion of score General principle : Occam razor Pluralitas non est ponenda sine neccesitate (La pluralité (des notions) ne devrait pas être posée sans nécessité) plurality should not be posited without necessity Frustra fit per plura quod potest fieri per pauciora (C est en vain que l on fait avec plusieurs ce que l on peut faire avec un petit nombre) It is pointless to do with more what can be done with fewer = Parcimony principle : find a model Fitting the data D : likelihood : L(D θ, B) The simplest possible : dimension of B : Dim(B) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 7/49

58 Score examples AIC and BIC Compromise between likelihood and complexity Application of AIC (Akaïke 70) and BIC (Schwartz 78) criteria S AIC (B, D) = log L(D θ MV, B) Dim(B) S BIC (B, D) = log L(D θ MV, B) 1 Dim(B) log N 2 Bayesian scores : BD, BDe, BDeu S BD (B, D) = P(B, D) (Cooper et Herskovits 92) BDe = BD + score equivalence (Heckerman 94) S BD (B, D) = P(B) n q i i=1 j=1 Γ(α ij ) Γ(N ij + α ij ) r i k=1 Γ(N ijk + α ijk ) Γ(α ijk ) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 8/49

59 Score properties Two important properties Decomposability (Global)Score(B, D) = n (local)score(x i, pa i ) i=1 Score equivalence If two BN B 1 and B 2 are Markov equivalent then S(B 1, D) = S(B 2, D) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 9/49

60 Score properties Two important properties Decomposability (Global)Score(B, D) = n (local)score(x i, pa i ) i=1 Score equivalence If two BN B 1 and B 2 are Markov equivalent then S(B 1, D) = S(B 2, D) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 9/49

61 Heuristic exploration of search space Search space and heuristics B space restriction to tree space : Chow&Liu, MWST DAG with node ordering : K2 algorithm greedy search genetic algorithms,... E space greedy equivalence search Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 10/49

62 Restriction to tree space Principle What is the best tree connecting all the nodes, i.e. maximizing a weight defined for each possible edge? Answer : maximal weighted spanning tree (MWST) weight = mutual information [Chow & Liu, 1968] W (X A, X B ) = a,b N ab N log N abn N a. N.b weight = any local score variation [Heckerman, 1994] W (X A, X B ) = score(x A, Pa(X A ) = X B ) score(x A, ) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 11/49

63 Restriction to tree space Principle What is the best tree connecting all the nodes, i.e. maximizing a weight defined for each possible edge? Answer : maximal weighted spanning tree (MWST) weight = mutual information [Chow & Liu, 1968] W (X A, X B ) = a,b N ab N log N abn N a. N.b weight = any local score variation [Heckerman, 1994] W (X A, X B ) = score(x A, Pa(X A ) = X B ) score(x A, ) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 11/49

64 Restriction to tree space Remarks MWST returns an undirected tree This undirected tree = CPDAG of all the directed tree with this skeleton Obtain a directed tree by (randomly) choosing one root and orienting the edges with a depth first search over this tree Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 12/49

65 Example : obtained DAG vs. target one A S A S T L B T L B O O X D X D MWST can not discover cycles neither V-structures (tree space!) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 13/49

66 Heuristic exploration of search space Search space and heuristics B space restriction to tree space : Chow&Liu, MWST DAG with node ordering : K2 algorithm greedy search genetic algorithms,... E space greedy equivalence search Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 14/49

67 Greedy search Principle Exploration of the search space with traversal operators add edge invert edge delete edge and respect the DAG definition (no cycle) Exploration can begin from any given DAG Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 15/49

68 Example : obtained DAG vs. target one A S A S T L B T L B O O X D X D start = empty graph. GS result = local optimum :-( Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 16/49

69 Example : obtained DAG vs. target one A S A S T L B T L B O O X D X D start = MWST result. GS result is better Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 16/49

70 Heuristic exploration of search space Search space and heuristics B space restriction to tree space : Chow&Liu, MWST DAG with node ordering : K2 algorithm greedy search genetic algorithms,... E space greedy equivalence search Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 17/49

71 What about changing our search space Preliminaries IC/PC result = CPDAG MWST result = CPDAG Score-based methods do not distinguish equivalent DAGs Search in E E = CPDAG space Better properties : YES 2 equivalent structures = 1 unique structure in E Better size : NO E size is quasi similar to DAG space asymptotic ratio is 3,7 : [Gillispie & Perlman, 2001] Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 18/49

72 Greedy Equivalent Search Principe [Chickering, 2002] Greedy search in E Phase 1 : add edges until convergence Phase 2 : delete edges until convergence Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 19/49

73 Add edge examples in E neighborhood of E 0 neighborhood of E 1 X 1 X 2 X 1 X 2 X 1 X 2 X 1 X 2 X 1 X 2 X 3 X 4 X 3 X 4 X 3 X 4 X 3 X 4 X 3 X 4 X 1 X 2 X 1 X 2 X 1 X 2 X 1 X 2 X 1 X 2 X 1 X 2 E 0 X 3 X 4 X 3 X 4 X 3 X 4 X 3 X 4 X 3 X 4 X 3 X 4 X 1 X 2 X 1 X 2 X 1 X 2 X 1 X 2 X 1 X 2 E 1 X 3 X 4 X 3 X 4 X 3 X 4 X 3 X 4 X 3 X 4 Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 20/49

74 Score-based methods How to search a good BN? Constraint-based methods BN = independence model find CI in data in order to build the DAG Score-based methods BN = probabilistic model that must fit data as well as possible search the DAG space in order to maximize a scoring/fitness function Hybrid methods Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 21/49

75 Hybrid methods = local search methods Local search and global learning Search one local neighborhood for a given node T Reiterate for each T Learn the global structure with these local informations which neighborhood? PC(T ) : Parents and Children T (without distinction) MB(T ) : Markov Blanket of T - Parents, children and spouses Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 22/49

76 Hybrid methods = local search methods Local search and global learning Search one local neighborhood for a given node T Reiterate for each T Learn the global structure with these local informations which neighborhood? PC(T ) : Parents and Children T (without distinction) MB(T ) : Markov Blanket of T - Parents, children and spouses Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 22/49

77 Local search identification Identification of MB(T) or PC(T) IAMB [Aliferis et al., 2002] MMPC [Tsamardinos et al., 2003],... Hybrid structure learning algorithms MMHC [Tsamardinos et al., 2006] = MMPC + Greedy search back Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 23/49

78 Outline... 4 Bayesian network Structure Learning constraint-based methods score-based methods local search methods 5 Causal Bayesian Networks definition causal BN structure learning 6 SemCaDo algorithm definition experimental study 7 O2C algorithm definition algorithm Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 24/49

79 A BN is not a causal model Usual BN A B does not imply direct causal relationship between A and B, Only edges from the CPDAG represent causal relationships Confusion When the DAG is given by an expert, this graph is very often causal When the DAG is learnt from data, no reason to be causal! Causal BN Each A B represents one direct causal relationship, i.e. A is the direct cause which generates B Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 25/49

80 A BN is not a causal model Usual BN A B does not imply direct causal relationship between A and B, Only edges from the CPDAG represent causal relationships Confusion When the DAG is given by an expert, this graph is very often causal When the DAG is learnt from data, no reason to be causal! Causal BN Each A B represents one direct causal relationship, i.e. A is the direct cause which generates B Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 25/49

81 A BN is not a causal model Usual BN A B does not imply direct causal relationship between A and B, Only edges from the CPDAG represent causal relationships Confusion When the DAG is given by an expert, this graph is very often causal When the DAG is learnt from data, no reason to be causal! Causal BN Each A B represents one direct causal relationship, i.e. A is the direct cause which generates B Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 25/49

82 Intervention vs. Observation Probabilistic inference : we observe B = b, we compute P(A B = b) Causal inference [Pearl 00]: we manipulate/intervene upon B : do(b = b) example with A B P(A do(b = b)) = P(A), P(B do(a = a)) = P(B A = a) example with A B P(A do(b = b)) = P(A B = b), P(B do(a = a)) = P(B) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 26/49

83 Causal structure learning Usual situation : observational data whatever the method, the right result is the CPDAG partial determination of the causal structure How to find a full causal graph? Use only experimental data, and decide at every step the more interesting experiment to realize (active learning [Murphy, 2001],...) Use only observational data, for a very specific distribution (LiNGAM models [Hoyer et al., 2008]) Another idea: MyCaDo algorithm [Meganck et al., 2006] Use (already existing) observational data to find the CPDAG Complete the orientation with experimental data Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 27/49

84 Causal structure learning Usual situation : observational data whatever the method, the right result is the CPDAG partial determination of the causal structure How to find a full causal graph? Use only experimental data, and decide at every step the more interesting experiment to realize (active learning [Murphy, 2001],...) Use only observational data, for a very specific distribution (LiNGAM models [Hoyer et al., 2008]) Another idea: MyCaDo algorithm [Meganck et al., 2006] Use (already existing) observational data to find the CPDAG Complete the orientation with experimental data Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 27/49

85 Causal structure learning Usual situation : observational data whatever the method, the right result is the CPDAG partial determination of the causal structure How to find a full causal graph? Use only experimental data, and decide at every step the more interesting experiment to realize (active learning [Murphy, 2001],...) Use only observational data, for a very specific distribution (LiNGAM models [Hoyer et al., 2008]) Another idea: MyCaDo algorithm [Meganck et al., 2006] Use (already existing) observational data to find the CPDAG Complete the orientation with experimental data Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 27/49

86 MyCaDo algorithm Données d'observation Algorithme d'apprentissage de structure d'un RB Système Données expérimentales CPDAG Choix de l'exp. Réalisation de l'exp. Analyse des résultats Réseau Bayésien causal Algorithme MyCaDo (1) Choice of the experiment = what variable M manipulate? the one potentially orienting more edges by taking into account experiment/observation cost back Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 28/49

87 MyCaDo algorithm Données d'observation Algorithme d'apprentissage de structure d'un RB Système Données expérimentales CPDAG Choix de l'exp. Réalisation de l'exp. Analyse des résultats Réseau Bayésien causal (2) Experimentation Algorithme MyCaDo do(m = m) for all possible values m observe all candidate variables C (C M) back Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 28/49

88 MyCaDo algorithm Données d'observation Algorithme d'apprentissage de structure d'un RB Système Données expérimentales CPDAG Choix de l'exp. Réalisation de l'exp. Analyse des résultats Réseau Bayésien causal Algorithme MyCaDo (3) Result analysis : P(C M) (obs.) P(C do(m)) (exp.)? if equal C M else M C orient some other edges by applying specific rules back Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 28/49

89 Outline... 4 Bayesian network Structure Learning constraint-based methods score-based methods local search methods 5 Causal Bayesian Networks definition causal BN structure learning 6 SemCaDo algorithm definition experimental study 7 O2C algorithm definition algorithm Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 29/49

90 SemCaDo assumptions BN Ontology general assumptions Nodes Concepts Random variables Concept attributes Causal dependencies Semantic causal relations Data Concept-attribute instances SemCaDo specific assumptions Causal relations concern concepts sharing the same semantic type Ontology continuity (evolution) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 30/49

91 SemCaDo assumptions BN Ontology general assumptions Nodes Concepts Random variables Concept attributes Causal dependencies Semantic causal relations Data Concept-attribute instances SemCaDo specific assumptions Causal relations concern concepts sharing the same semantic type Ontology continuity (evolution) Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 30/49

92 Example Catastrophes is-a is-a Man-made Natural is-a is-a is-a is-a is-a is-a Flood Fire Earthquake Landslide Volcano Tsunami Causes Causes Causes Causes Causes Causes Our causal BN will represent the grey part of the ontology Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 31/49

93 SemCaDo three main steps Domain ontology Inter. Data 1 st step Structure learning Choice of experience 3 rd step Ontology evolution Perform the experience Obs. Data PDAG Analyse the results CBN 2 nd step Causal discovery Enriched ontology Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 32/49

94 1 st step: initial BN structure learning Extraction of the causal relationships (R ) from the ontology Integration of these edge as constraints in the structure learning algorithm [De Campos & al., 2007] Continuity : these edges will not be questioned during learning Interest Ontology helps in reducing the search task complexity Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 33/49

95 1 st step: initial BN structure learning Extraction of the causal relationships (R ) from the ontology Integration of these edge as constraints in the structure learning algorithm [De Campos & al., 2007] Continuity : these edges will not be questioned during learning Interest Ontology helps in reducing the search task complexity Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 33/49

96 2 nd step: serendipitous causal discovery Experimentations are needed in order to find the causal orientation of some edges Our solution : MyCaDo [Meganck & al., 2006], iterative causal discovery process Adaptation to take into account ontological knowledge : Rada distance on H between one set of concepts and their most specific common subsumer Interest Ontology helps in potentially orienting the more unexpected (serendipitous) links Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 34/49

97 2 nd step: serendipitous causal discovery Experimentations are needed in order to find the causal orientation of some edges Our solution : MyCaDo [Meganck & al., 2006], iterative causal discovery process Adaptation to take into account ontological knowledge : Rada distance on H between one set of concepts and their most specific common subsumer Interest Ontology helps in potentially orienting the more unexpected (serendipitous) links Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 34/49

98 3 rd Step: Ontology evolution process [Stojanovic et al., 2002] Interest BN structure learning from data helps in discovering new relations in the ontology Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 35/49

99 3 rd Step: Ontology evolution process [Stojanovic et al., 2002] Interest BN structure learning from data helps in discovering new relations in the ontology Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 35/49

100 Experimental study (1) Benchmark : no existing real benchmark & system :-( BN graph : random generation (50 to 200 nodes) Ontology : Causal relationships : BN edges Hierarchy of concept : generation by clustering BN nodes Data is generated by using BN as a generative model Experimental protocol Hierarchy of concepts and 10% to 40% of existing causal relationships are given as inputs Semantic gain : cumulative Rada distance of the discovered relationships Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 36/49

101 Experimental study (1) Benchmark : no existing real benchmark & system :-( BN graph : random generation (50 to 200 nodes) Ontology : Causal relationships : BN edges Hierarchy of concept : generation by clustering BN nodes Data is generated by using BN as a generative model Experimental protocol Hierarchy of concepts and 10% to 40% of existing causal relationships are given as inputs Semantic gain : cumulative Rada distance of the discovered relationships Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 36/49

102 Experimental study (2) Ontology helps structure learning SemCaDo performs better in less steps than MyCaDo back Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 37/49

103 Experimental study (2) Structure learning helps ontology evolution Original causal discoveries are discovered first and can be added to the ontology back Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 37/49

104 Outline... 4 Bayesian network Structure Learning constraint-based methods score-based methods local search methods 5 Causal Bayesian Networks definition causal BN structure learning 6 SemCaDo algorithm definition experimental study 7 O2C algorithm definition algorithm Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 38/49

105 Object Oriented Bayesian Networks An extension of BNs using the object paradigm [Bangsø and Wuillemin, 2000a; Bangsø and Wuillemin, 2000b; Koller and Pfeffer, 1997] Support several aspects of the object oriented modeling. (e.g., inheritance, instantiation) Designed to model large and complex domains OOBN structure learning : OO-SEM This algorithm [Langseth and Nielsen, 2003] is based on 2 steps Generation of a prior OOBN based on a prior expert knowledge Grouping nodes into instantiations and instantiations into classes Giving a prior information about the candidate interfaces Adaptation of Structural EM algorithm to learn the final structure Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 39/49

106 Object Oriented Bayesian Networks An extension of BNs using the object paradigm [Bangsø and Wuillemin, 2000a; Bangsø and Wuillemin, 2000b; Koller and Pfeffer, 1997] Support several aspects of the object oriented modeling. (e.g., inheritance, instantiation) Designed to model large and complex domains OOBN structure learning : OO-SEM This algorithm [Langseth and Nielsen, 2003] is based on 2 steps Generation of a prior OOBN based on a prior expert knowledge Grouping nodes into instantiations and instantiations into classes Giving a prior information about the candidate interfaces Adaptation of Structural EM algorithm to learn the final structure Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 39/49

107 OOBN structure learning Drawbacks This prior knowledge is not always obvious to obtain The expert should be familiar with the object oriented modeling Our idea Harness ontologies representation capabilities in order to generate the prior OOBN structure Ontologies Concepts Cp Properties P cpi Inheritance relations H R Semantic relations S R OOBNs Classes Real nodes Class hierarchies Links/ Interfaces Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 40/49

108 OOBN structure learning Drawbacks This prior knowledge is not always obvious to obtain The expert should be familiar with the object oriented modeling Our idea Harness ontologies representation capabilities in order to generate the prior OOBN structure Ontologies Concepts Cp Properties P cpi Inheritance relations H R Semantic relations S R OOBNs Classes Real nodes Class hierarchies Links/ Interfaces Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 40/49

109 The 2OC approach Ontology Morphing process Prior OOBN Learning process Data (1) ontology to prior OOBN [Ben Ishak et al., 2011a] Final OOBN Change detection process Proposals of evolution it seems interesting to refine the ontology used initially! (2) final OOBN to ontology [Ben Ishak et al., 2011b] Montassar Ben Messaoud, Mouna Ben Ishak, Philippe Leray, Nahla Ben Apprentissage Amor RB et Ontologies 41/49