Multi-source RSC model for multiple search result clustering

Multimedia search engines are often based on multiple decentralized search services, multiple information sources (text search, audio search, visual search, semantic search engines, etc.), multiple data representation and similarity measures. Heterogeneous multiple search results need to be combined and structured efficiently and generically. In this paper, we propose a new multiple search results clustering algorithm based on the Relevant Set Correlation model, it constitutes an extension of the previously proposed algorithm. As the model is based on shared neighborhood information only, it allows our new technique to process the different information sources as simple oracles returning ranked lists of relevant objects. We experimented the proposed clustering framework on two image datasets with 5 different CBIR techniques as oracles. The results show that our new multi-source scheme performs better than using each source independently and that it is robust to the presence of noisy oracles. We also show that it performs better than using an early fusion strategy although this fusion approach makes use of more information about the used visual features and similarity measures.