Research On User Interest Mining And Context-aware Recommendation System Based On Topic Model

With the development of Internet technology,digital resources grow exponentially.In the field of smart TV,every day a massive amount of video data is generated.User interaction is diverse,and the number of user behavior begin to surge.With the national "triple play" and other strategies to promote,and smart TV users continue to expand.How to deal with and effectively use large-scale data has become an urgent problem in the field.Because of the big data,search engines,personalized recommendation technology,this problem has been taken seriously,and gradually derived from some solutions.At present,the personalized recommendation system has been widely studied and applied.It can help users to better tap their own interests,and help build the user profile of the system.It helps maintain the user's attention,and help avoid the loss of users.The main methods of the recommended system are model-based approaches and memory-based approaches.Model-based recommendation system can accurately express the user's interest,and outperforms the other methods at most time.Memory-based approaches are relatively simple and easy,and has good interpretability.How to effectively combine the strengths of these two models is a research focus of this thesis.In addition,in the smart-TV recommendation system,because the TV is a shared client,the user's interest in different time will be quite different.How to introduce the time-aware concept to the model is another focus of this article.Therefore,in this thesis,we first propose a recommendation algorithm based on short text LDA model.This algorithm is a semantic-based method.It applies the latent semantic model of the text mining field to the recommendation system to accurately construct the user's interest.In order to address the sparsity issue,we directly model the video-pair of one user instead of single video.This method greatly solves the sparsity problem and can effectively improve the accuracy of user interests.We introduce the modified LDA model for shot text,and model the user viewing records into low-dimensional space which consists of two matrices,namely the user interest matrix(user-topic),and topic-video matrix.Then we take time information into consideration,and solve the share-account problem in smart-TV recommendation systems.The algorithm is a neighborhood-based recommendation algorithm,which recommends video to users with similar interests.In the process of constructing user interest,the context-aware recommendation strategy is used.We construct video-pairs with context constraints,and only those in the same context will be used.This pre-filtering strategy effectively uses time context information,which can distinguish different user interests in different time periods and also avoid constructing unrelated video into the same video-pair.In addition,when the recommendation list is retrieved,the context-aware context recommendation strategy is introduced once again.Whether each video is worth recommending in the current environment will be determined by the user's interest in the current context.This post-filtering method can further rearrange recommendation list according to the time context information,and greatly improve the recommendation effect.In order to test the effectiveness of the model,we use a well-known TV recommendation platform,Hisense TV cloud platform.We provide a variety of contrast recommendation algorithm,and a variety of evaluation measures.Our method has achieved higher recall rate,MAP and MRR than the other traditional recommendation algorithms,which proves the effectiveness of the proposed method.