Learn to Personalized Image Search from the Photo Sharing Websites

 ABSTRACT:

            Increasingly developed social sharing websites, like Flickr and 

Youtube, allow users to create, share, annotate and comment Medias. 

The large-scale user-generated meta-data not only facilitate users in 

sharing and organizing multimedia content, but provide useful 

information to improve media retrieval and management. Personalized 

search serves as one of such examples where the web search experience 

is improved by generating the returned list according to the modified 

user search intents. In this paper, we exploit the social annotations and 

propose a novel framework simultaneously considering the user and 

query relevance to learn to personalized image search. The basic 

premise is to embed the user preference and query-related search intent 

into user-specific topic spaces. Since the users’ original annotation is too 

sparse for topic modeling, we need to enrich users’ annotation pool 

before user specific topic spaces construction.

The proposed framework contains two components:

   1) A Ranking based Multi-correlation Tensor Factorization model is 

proposed to perform annotation prediction, which is considered as users’ 

potential annotations for the images.

        2) We introduce User-specific Topic Modeling to map the query relevance 

and user preference into the same user-specific topic space. For performance 

evaluation, two resources involved with users’ social activities are employed. 

Experiments on a large-scale Flickr dataset demonstrate the effectiveness of 

the proposed method.