Unsupervised clustering of clickthrough data for automatic annotation of multimedia content

Abstract

Current low-level feature-based CBIR methods do not provide meaningful results on non-annotated content. On the other hand manual annotation is both time/money consuming and user-dependent. To address these problems in this paper we present an automatic annotation approach by clustering, in an unsupervised way, clickthrough data of search engines. In particular the query-log and the log of links the users clicked on are analyzed in order to extract and assign keywords to selected content. Content annotation is also accelerated by a carousel-like methodology. The proposed approach is feasible even for large sets of queries and features and theoretical results are verified in a controlled experiment, which shows that the method can effectively annotate multimedia files