A Declarative Similarity Framework for Knowledge Intensive CBR
A Declarative Similarity Framework for Knowledge Intensive CBR by Díaz-Agudo and González-Calero Presented by Ida Sofie G Stenerud 25. October 2006
Terminology p CBROnto p Description Logic p LOOM
CBROnto p ”A task based ontology compromising common CBR terminology” p A vocabulary for expressing the CBR elements p CBROnto’s 2 purposes: n n To integrate CBR process knowledge and domain knowledge To be a domain-indepentent framework for designing CBR systems
Description Logics p A knowledge representation language p Formal objects: n n n p Concepts Relations Individuals Reasoning mechanisms n n Subsumption Instance Recognition
LOOM p A description logic implementation p A query language for CBR p Example:
The Similarity Framework p Several similarity measures can coexist at one time p Several approaches to case retrieval: n Relevance Criteria p n Similarity Criteria p n Declarative represetation of differences and similarity Representational approach p n User-defined or pre-defined in program ”Semantic traversal” of the hierarchy Computational approach p Can be a combination of all the above
The Similarity Framework
User-defined Relevance Criteria p Relevance: Why is this case more relevant than other cases? p More-on-point, Most-on-point
Similarity Terms p Explicit computation of similarity terms p Declaratively p Similarity: ”the most specific concept which subsumes 2 cases”
Representational Approach p Assignment of similarity meanings to the path between cases p The Generic Travel Operator
Computational Approach p Different alternatives to compute numeric similarity between attributes n n Nearest Neighbour Algorithm The classic global similarity approach For common attributes: use local measure p For common relations: use global measure to compare related sub-objects p Similarity is a weighted sum of these p p NB: Only attribute level, not instance level similarity!
Intra-class vs. inter-class similarity p Intra-class similarity: n p Inter-class similarity n p Dependent on the attribute fillers Dependent on object position in hierarchy Multiplied to get final similarity result
The Similarity Framework
Similarity Functions p Local similarity n p Global similarity n p Similarity between two values of a type Defines how you combine local similarities Positional similarity n Indepentent of attribute values, only dependent on position in hierarchy
A Case Matching Example p Case 1 n n Color: Dark Green Price: 1 500 000 p Case 2 n n Color: Dark Gray Price: 1 100 000
Discussion Any questions?
- Slides: 16