The Collaborative Topic Regression Combined With Sequential Behaviors

With the emergence of information overload,it becomes more and more difficult for us to get information which we really need in the increasingly open Internet.The personalized recommendation can effectively solve the problem of information overload and recommends the information and products which users are interested in forwardly.Collaborative filtering is the most widely used method in personalized recommendation,but since it generally just considers user-item ratings as the unique recommend data,there are problems such as cold start and sparseness.In this thesis,basing on the novel hierarchical Bayesian model—collaborative topic regression model(CTR)which combines the topic model(LDA)with the probability matrix factorization model(PMF),we use the topic model to tackle the tag information of item and modify the probability matrix factorization model respectively.We not only take the user-item ratings into account,but also add the user's trust relationship,sequential behaviors,tag information of item which have effect on the recommendation to the model.According to the recommendation of their friends and trusted users,users can choose what they are interested in.The relationship of users based on sequential behaviors also has an impact on the user's choice.The user latent feature vector is generated by the PMF model which is modified by means of linear fusing the influence of the sequential behaviors on the user relationship and the trust value.In addition,the tag information defined by the users can also reflect the preference of users to a certain extent.Therefore,we use the topic model LDA to process the tag information of the item to get the item latent feature vector.Finally,according to the basic principle of CTR model,we propose N-CTR model which is mixed the characteristics of improved LDA model and PMF model into the CTR.We optimize the final user latent feature matrix,item latent feature matrix and theme distribution vector which are utilized to predict ratings by gradient decent method and expectation maximization algorithm.The experiments are carried out on the dataset of Last.fin.Compared with PMF model which only considers user-item ratings,the result of experiment shows that the recommendation accuracy rates of the N-CTR model,which takes many factors such as user's trust relationship,sequential behaviors,tag information of item into account,are improved by 7.36%and 7.94%respectively in MAE and RMSE.It indicates that N-CTR model not only alleviates sparsity problem in the process of recommendation,but also gets more precise recommendation accuracy than the traditional model-based collaborative filtering.