Improving Image Search based on User Created Communities
Tag-based retrieval of multimedia content is a difficult problem, not only because of the shorter length of tags associated with images and videos, but also due to mismatch in the terminologies used by searcher and content creator. To alleviate this problem, we propose a simple concept-driven probabilistic model for improving text-based rich-media search. While our approach is similar to existing topic-based retrieval and cluster-based language modeling work, there are two important differences: (1) our proposed model considers not only the query-generation likelihood from cluster, but explicitly accounts for the overall "popularity" of the cluster or underlying concept, and (2) we explore the possibility of inferring the likely concept relevant to a rich-media content through the user-created communities that the content belongs to. We implement two methods of concept extraction: a traditional cluster based approach, and the proposed community based approach. We evaluate these two techniques for how effectively they capture the intended meaning of a term from the content creator and searcher, and their overall value in improving image search. Our results show that concept-driven search, though simple, clearly outperforms plain search. Among the two techniques for concept-driven search, community-based approach is more successful, as the concepts generated from user communities are found to be more intuitive and appealing.