Learning compact hashing codes with complex objectives from multiple sources for large scale similarity search

Similarity search is a key problem in many real world applications including image and text retrieval, content reuse detection and collaborative filtering. The purpose of similarity search is to identify similar data examples given a query example. Due to the explosive growth of the Internet, a huge amount of data such as texts, images and videos has been generated, which indicates that efficient large scale similarity search becomes more important. Hashing methods have become popular for large scale similarity search due to their computational and memory efficiency. These hashing methods design compact binary codes to represent data examples so that similar examples are mapped into similar codes. This dissertation addresses five major problems for utilizing supervised information from multiple sources in hashing with respect to different objectives. Firstly, we address the problem of incorporating semantic tags by modeling the latent correlations between tags and data examples. More precisely, the hashing codes are learned in a unified semi-supervised framework by simultaneously preserving the similarities between data examples and ensuring the tag consistency via a latent factor model. Secondly, we solve the missing data problem by latent subspace learning from multiple sources. The hashing codes are learned by enforcing the data consistency among different sources. Thirdly, we address the problem of hashing on structured data by graph learning. A weighted graph is constructed based on the structured knowledge from the data. The hashing codes are then learned by preserving the graph similarities. Fourthly, we address the problem of learning high ranking quality hashing codes by utilizing the relevance judgments from users. The hashing code/function is learned via optimizing a commonly used non-smooth non-convex ranking measure, NDCG. Finally, we deal with the problem of insufficient supervision by active learning. We propose to actively select the most informative data examples and tags in a joint manner based on the selection criteria that both the data examples and tags should be most uncertain and dissimilar with each other. Extensive experiments on several large scale datasets demonstrate the superior performance of the proposed approaches over several state-of-the-art hashing methods from different perspectives