Prior-based probabilistic latent semantic analysis for multimedia retrieval

Abstract

Topic models have shown to be one of the most effective tools in Content-Based Multimedia Retrieval (CBMR). However, the high computational learning cost together with the huge expansion of multimedia collections limit the scalability of topic-based CBMR systems in real-life multimedia applications. The present work pursues a twofold objective. On the one hand, to study the effect of using clustering-based document reduction schemes over standard topic models pLSA (probabilistic Latent Semantic Analysis) and LDA (Latent Dirichlet Allocation). On the other hand, to develop a pLSA-based extension oriented to integrate this reduction scheme within the own model in order to improve the CBMR effectiveness. The experimental part of the work includes three different multimedia databases, three ranking functions, four retrieval scenarios, three different numbers of topics and ten document reduction levels. Experiments revealed that standard topic models are highly sensitive to the document reduction level whereas the proposed model is able to provide a competitive advantage within the content-based retrieval field.