An efficient solution to factor drifting problem in the pLSA model

Probabilistic Latent Semantic Analysis (pLSA) is a powerful statistical technique to analyze relation between factors in dyadic data Although various pLSA-based applications, ranging from information retrieval, information filtering, to text-mining and visualization, have been successfully conducted, they can not afford dynamic revising of model when one of the factors changes constantly. In this paper, we take the advantage of decoupling ability of pLSA thoroughly, and propose a more elegant approach based on maximum likelihood estimation to gain an incremental learning with the drift of a factor. We demonstrate our method in the context of collaborative filtering where single user interests change fast, but the community interests remain almost constant. Experiments against the MovieLens and EachMovie data sets reveal that the proposed method improves the recommending accuracy 10% further beyond the original pLSA at a less computation cost.