Replacing log-based profiles to context profiles and its application to context-aware document clustering

As the number of documents is increasing rapidly, personalization is becoming more and more important to find desired documents efficiently. In typical personalization frameworks, a user profile created by its histories is necessary but it may include different contexts even for one user. In this paper, we develop a framework of personalization for information retrieval in various contexts. The main idea of this paper is twofold: firstly we use a vector as a generalized profile, called a context profile, of a user, a context, or a segment, and secondly we use corpora instead of user histories. This means that we can create a profile for a context at low cost and choose it according to contexts. Moreover, we can easily obtain virtual profiles from created profiles since profiles are just vectors. To evaluate the proposed framework, we have created many context profiles from a popular corpus, adjusted usual document vectors to contexts, and compared to adjusted document vectors and original ones. Effectiveness of adjusted documents are also confirmed by document clustering which creates different clusters according to contexts.