Case-Based Solution Diversity Alexandra Coman Héctor Muñoz-Avila Dept. of Computer Science & Engineering Lehigh University Sources: cbrwiki.fdi.ucm.es/

Outline Lehigh University The InSyTe Laboratory Overview of Case-Based Reasoning  Similarity  Retrieval  Adaptation Conversational Case-based reasoning Diversity versus Similarity General versus Episodic Knowledge Final Remarks

Synthetizing Diversity Showcasing diverse solutions: success story in recommender systems (Smyth, Burke, McGinty …) Plan diversity:  Definition of the problem: quantitative vs qualitiative (Myers, AAAI-01)  Generating two or more quantitative different plans for same problem (Srivastava et al, IJCAI-07) Synthetizing diversity:  Case-based retrieval and adaptation from plan library (Coman & Munoz-Avila, ICCBR-10; 11 – under review )  Generating two or more qualitatively different plans for same problem (Coman & Munoz-Avila, AAAI-11)  Our common solution: S: diverse solutions so far, s: candidate solution, P: new problem sim(s,P) + relativeDiversity(s,S) What changes: S, s, P, sim(), D(s,s’) 11

Research Program: Synthetizing Diversity Plan DiversityCase-based plan diversity preliminary work: New insight: sim(s,P) + relativeDiversity(s,S) Proposed idea: sim(s,P) + relativeDiversity(s,S) + cost(s) Danger: don’t want it to be a planning proposal Research topics: Representation scope of using D() versus qualitative diversity Trade-offs of solution: Diversity versus quality Diversity versus generation Diversity in other paradigms: search (A*)

Focus Point: Diversity in CBR

Traditional Retrieval Approach  Similarity-Based Retrieval  Select the k most similar items to the current query.  Problem  Vague queries.  Limited coverage of search space in every cycle of the dialogue. C2 C1 C3 Q Query Available case Similar case Lorraine McGinty and Barry Smyth Department of Computer Science, University College Dublin

Diversity Enhancement Lorraine McGinty and Barry Smyth Department of Computer Science, University College Dublin  Diversity-Enhanced Retrieval  Select k items such that they are both similar to the current query but different from each other.  Providing a wider choice allows for broader coverage of the product space.  Allows many less relevant items to be eliminated. C2C3 C1 Q Query Available case Retrieved case

Dangers of Diversity Enhancement Lorraine McGinty and Barry Smyth Department of Computer Science, University College Dublin  Leap-Frogging the Target  Problems occur when the target product is rejected as a retrieval candidate on diversity grounds.   Protracted dialogs.  Diversity is problematic in the region of the target product.  Use similarity for fine-grained search.  Similarity is problematic when far from the target product.  Use diversity to speed-up the search. T C2C3 C1 Q

Final Remarks

Questions?