FTM: Recommending the Right Items for User Temporal Interests with Matrix Factorization through Topic Model

摘要

Historical user activity is the key for building user profiles to predict the user behavior and affinities in many web applications such as targeting of online advertising and social recommendations. In these scenarios, the items recommended to users must match their profiles. However, user profiles are temporal, so, changes in a user's activity patterns are particularly useful for improved prediction and recommendation. In our paper, we use matrix factorization method to predict the temporal preferences of users on items. Because of data sparseness, regularization is the key to good predictive accuracy. Our method works by regularizing both user and item factors simultaneously through user dynamic interests and the topics associated with each item. Specifically, to regularize user, we use dynamic topic model to model the temporal interest associated with each user. All the historical activities on the web of a user are considered as a document and each activity is a word. Analogous to a word belongs to a topic, each activity is also associated with a "behavioral tendency", user behavioral tendencies are obtained by using dynamic topic model. Our method models behavioral tendencies of a user dynamically where both the user association with the behavioral tendencies and the behavioral tendencies themselves are allowed to vary over time, thus ensuring that the profiles remain current. To regularize an item, we treat each word in the item is associated with a discrete latent factor often referred to as the topic of the word, item topics are obtained by averaging topics across all words in an item. Then, user preference on an item is modeled as the affinity of user behavioral tendencies to the item's topics. Additionally, to better model a user preference on an item, we also consider the user historical preferences on other items, called "user bias". The item popularity is also a factor that affects the user preference on this item. In a word, we incorporate all the above fourth factors (user temporal interest, item topics, user bias, item popularity) into the matrix factorization framework, and proposed a new uniform approach. This approach can not only model user's dynamic interest, but also predict the preference of a user on an item.