期刊名称：International Journal of Computer Science & Technology
印刷版ISSN：2229-4333
电子版ISSN：0976-8491
出版年度：2013
卷号：4
期号：4
页码：428-437
语种：English
出版社：Ayushmaan Technologies
摘要：Increasingly developed social sharing websites, like Flickr andYoutube, allow users to create, share, annotate and commentmedias. The large-scale user-generated meta-data not only facilitateusers in sharing and organizing multimedia content, but provideuseful information to improve media retrieval and management.Personalized search serves as one of such examples where theweb search experience is improved by generating the returnedlist according to the modifed user search intents. In this paper,we exploit the social annotations and propose a novel frameworksimultaneously considering the user and query relevance to learnto personalized image search. The basic premise is to embed theuser preference and query-related search intent into user-specifctopic spaces. Since the users’ original annotation is too sparse fortopic modeling, we need to enrich users’ annotation pool beforeuser-specifc topic spaces construction. The proposed frameworkcontains two components: (1) A Ranking based Multi-correlationTensor Factorization model is proposed to perform annotationprediction, which is considered as users’ potential annotationsfor the images; (2) We introduce User-specifc Topic Modeling tomap the query relevance and user preference into the same userspecifc topic space. For performance evaluation, two resourcesinvolved with users’ social activities are employed. Experimentson a large-scale Flickr dataset demonstrate the effectiveness ofthe proposed method.
关键词：Personalized Image Search;Tensor Factorization;Topic Model; Social Annotation
