The Context-aware Recommendation Based On Tensor Factorization And Application

Recommender system is one of the current popular Internet applied systems, includesthe specialized sofeware tools and techniques for advising the user what products to select.Recommender system is the sub-discipline of information filtering technology. When get-ting large number of available information, the system applies the data mining techniquesand predicted algorithms to predict the ratings on the specific users to items.Traditional rec-ommender systems usually use the item-based or user-based recommender algorithm, thecollaborative filtering based item-similarity or user-similarity and the hybrid recommenderalgorithms combined both the algorithms in practical applications.The core of recommender system is the recommender algorithm.The main applied al-gorithms include the collaborative filtering recommender algorithm based item-similarity oruser-similarity and the data mining recommender algorithm based association rules or clus-tering. When handling the user-item binary relation, the Matrix Factorization has betterperformance. With the emerging of social tags, the research on user-item-tag ternary relationis necessary. For ternary relations research, the more advanced recommender algorithm isthe Tensor Factorization recommendation model, such as the CP-decomposition, the HighOrder Singular Value Decomposition and the Multiverse Recommendation.Our main work has done as follows:(1)First, we elaborate the recommender systems and their development in detail.Wegeneralize the classification of recommender systems and their related applications. More-over we note that the main problems of recommendation systems facing in development. (2)Then we introduce the concept of tensor,the matrix factorization which handls bina-ry relation and and the tensor factorization which handls diverse relationships. Specifically, we give the different concept of tensor in physics, algebra and computer application. We introduce the matrix factorization theory and give examples to show how matrix factoriza-tion apply in music recommendation; The Tensor Factorization is the higher dimensional promotion of Matrix Factorization. The tensor can make good use of the "user-item-context" ternary relationship to construct tensor model to implement recommendation.We describe two algorithms of tensor factorization, the CP-tensor decomposition mod-el and high Singular value decomposition model.(3)Finally, the forth chapter gives our three-dimensional and high-dimensional tensor factorization model with the context variable. We add the implicit feedback information to the "user-item-context" ternary construction as the third dimension of model.In this way we improve the High Order Singular Value Decomposition to multi-dimensional.The objective function: can be minimized to obtain the factor matrices and the core tensor G*,Then take the factor matrices and core tensor into the approximate By making use of the formula we can estimate user preferences for all products, and then make further optimization.Ultimately we give the users the optimal product list Top-N in a specific context implicit action.In particular online shopping behavior, we collect real data in the dataset TMall. And we point out how the user’s actions such as click、collect、cart、alipay and visit time are quantified into context variables. Then we apply the tensor factorizationmodel to recommend.(4)In experimental part, we use the TMall dataset and other datasets to verify the ef-fectiveness of algorithm. We preprocess the raw dataset to suitable format.We use the ex-perimental result show that the CCTF algorithm greatly improve on hit ratio comparing withthe matrix factorization. Under diferent levels of the context influence,we analysis theefciency of four similar algorithms.