Improving Similarity Search Ranking Outline Similarity Search Definition
- Slides: 6
Improving Similarity Search: Ranking
Outline § § § Similarity Search § Definition, state-of-the-art, challenges Ranking § Principle, use in similarity searching § Ranking strategies User Satisfaction Experiment § Description, demo
Similarity search § Basic principle § Describe data objects by their important properties (e. g. extract visual descriptors from images) § Using such description, search for most similar objects § State-of-the-art § Working search engines (MUFIN) § Results not always satisfactory § Difficulties § Basic image descriptors do not enable precise search § More precise descriptors need too much space and time § Moreover, similarity is subjective and context-dependent
Ranking § Basic task § Given a set of objects, order them in such way that the most interesting ones come first § We expect the set to be relatively small, therefore more complicated evaluation of similarity can be used § Search scenario using ranking § § Use some method for searching, obtain initial results Use ranking to improve the results Google solution: text search + similarity ranking MUFIN solution: similarity search + similarity ranking
Ranking Strategies § Predefined ranking § § Same for all results, chosen in advance Possible implementation: other similarity measure (local descriptors, text annotations, …) § User-defined ranking § § User can influence the ranking according to his/her preferences Possible implementation: relevance feedback principle, user chooses relevant objects from initial result, these are used for ranking
User Satisfaction Experiment § Ranking based on user feedback http: //mufin. fi. muni. cz/~xkohout 7/satisfaction. php
