Eswc2009
- 1. ReduCE: A Reduced Coulomb Energy Network Method for Approximate Classification Dipartimento di Informatica Università degli studi di Bari Nicola Fanizzi Claudia d'Amato Floriana Esposito
- 2. Table of Contents Motivation
- 3. Learning RCE Networks
- 4. Approximate Classifications of Individuals
- 5. Experiments
- 7. Motivation Classic ML techniques for building inductive classifiers for SW representations explicit models: new concepts
- 8. implicit models: neural networks, support vector machines, graphic probabilistic models Inductive methods for classification often more efficient and noise-tolerant than standard methods
- 10. better exploitation of the inherently incomplete information in Kbs for specific tasks
- 11. Applications of Inductive Models Approximate instance-checking
- 12. This can be also exploited for approximate retrieval
- 13. subsumption
- 14. ... It also provides alternative methods for ontology population
- 15. Ultimately, may be used for completing ontologies with probabilistic assertions enabling more sophisticate approaches to dealing with uncertainty
- 16. Learning Problem Given a target concept
- 17. Train a model (hypothesis) h Q using: Set of pre-classified individuals: examples
- 18. A knowledge base K as background knowledge then
- 19. Use the learned model to classify all other individuals: Given h Q and x 0
- 20. Output h Q ( x 0 ) and possibly the likelihood of this assertion
- 21. Training Set A limited number of individuals for which the intended classification is known Hypothesis: the function to be approximated
- 22. The Inductive Model: RCE Networks category layer a cj pattern (radii) layer w jk input layer . . . . . .
- 24. RCE Model Construction 1 2 3 4 5 6
- 25. Similarity Measure Generalization of a pseudo-distance d [ESWC2008]
- 26. RCE Final Model Prototypes (examples)
- 27. ( Vanilla ) Classification Procedure
- 28. Extensions generalizing the decision-making step:
- 29. likelihood:
- 30. Experiments For each ontology 100 satisfiable query concepts randomly
- 31. generated by composition (conjunction / disjunction) of (2 through 8) primitive and defined concepts Evaluation comparing inductive responses to those returned by a standard reasoner (Pellet 2) Indices match rate : identical classification
- 32. omission error rate : 0 vs. ± 1
- 33. commission error rate : +1 vs. -1 or -1 vs. +1
- 34. induction rate : ± 1 vs. 0 Several runs: rates averaged according to the 632+ bootstrap procedure
- 35. Experiments: Ontologies Ontologies employed in the experiments:
- 36. Outcomes
- 37. Outcomes / 2
- 38. Outcomes / 3
- 39. Conclusions & Outlook ML method transposed to SeWeb representations
- 40. Experiments: good performance High match rate
- 42. Some omissions
- 44. Low variance wrt to past inductive methods Extensions force binary response Expected to augment induction rate Pre-computation of prototypical individuals: Medoids from Clustering Use likelihood for ranking
- 45. adding probabilities to assertions
- 46. Questions ? For offline contacts: Nicola Fanizzi [email_address]
- 47. Claudia d'Amato [email_address]
- 48. Floriana Esposito [email_address] Other methods / systems http://lacam.di.uniba.it:8000/~nico/research/ontologymining.html The End