A Research Of Modeling And Predicting For User Preferences And Purchase Behavior

With the rapid growth of the Internet and the overwhelming amount of contents and choices that people are confronted with,recommender systems have been developed to facilitate the decision making process.During the past decade,the recommender system has proved to be a efficient way to handle data overload.It has been widely studied and applied both in academic and industry.Recommender system plays an important role in the internet services,greatly improving the efficiency of information retrieval and improving the quality of personalized services.Recommender system is mainly to provide personalized services to individuals or specific groups.In order to provide natural recommendation services,one must possess adequate knowlege of the users’ preferences as well as their behaviors pattern.Understanding user preference is essential to the optimization of recommender systems.As a feedback of user’s taste,rating scores can directly reflect the preference of a given user to a given product.Uncovering the latent components of user ratings is thus of significant importance for learning user interests.On the other hand,customer loyalty is crucial for internet services since retaining users of a service is of significance for increasing revenue.And customer loyalty can be reflected by how ofter a customer stays on or return to a website.In this paper,to better capture the users’ preference,a new recommendation approach was proposed by investigating the latent components of user ratings.We assumed that individual users’ evaluations on items are multi-criteria,and in a recommender system,their ratings consist of multiple latent components and were uncovered via a costsensitive learning strategy.Specifically,each rating is assigned to several latent factor models and the predictive errors for each rating with respect to those models is reserved.Then the accumulated predictive errors of models are utilized to decompose a rating into several independent components,which are utilized to further retrain the latent factor models.Finally,all the latent factor models are combined linearly to estimate predictive ratings for users.For modeling the users’ behaviors,we exploit the rich interaction data of users to build a customers retention evaluation model focusing on the return time of a user to a product,due to the length of the user’s stay reflects the viscosity of the recommendation service and its impact on the user.The model is established based on Cox’s proportional hazard that provides benefits for user-reserved dynamics and one can easily incorporate covariates into the model.We jointly model multiple aspects,namely the consilience between user and product,the sensitivity of the user to price and the external in uence the user might receive,via a probability approach,under the Cox’s proportional hazard framework.Finally,extensive experiments on real world data have demonstrated the superiority of our proposed model over state-of-the-art algorithms in both users’ preference modeling and users’ behaviors modeling.