Research On Recommender System Based On Matrix Factorization Technique With Latent Factors

The rapid development of internet technology not only promotes the rapid popularization of e-commerce but also bring us the technology of recommendation system which brings more convenience for our consumption.The Rating model based on matrix latent factor in collaborative filtering recommendation represents high-accuracy and low time complexity.The model consists of two parts: linear model and implicit factor model.However,the previous research focused on improving the accuracy under the circumstances of ensuring the efficiency during the calculation,as a result,the linear bias combination effect has not been deeply studied before.Now,we intend to analyze and improve this problem by considering the necessary time information.Our thesis will organize as follows:1.This work studies how the combination of priori bias and training bias influences the prediction accuracy of the model.This paper designs a comparison experiment by different combination of priori bias and training bias.The result of the experiment shows that bias could have a positive or negative effect on the model.While the global average bias of this parameter is able to improve the accuracy of the model stability.2.This work studies the matrix factorization model based on time information,and propose three new algorithms based on user drifting.The first is to improve the accuracy of the algorithm by discovering the relatively stable user drifting in a fixed time window,considering of the experience of modeling item drifting this algorithm believes that the user's interest drifting is also contain fixed part and can be extracted.The second method is to refine the linear model to capture user's long-term interest drifting.This paper measures user's long-term drifting by using the difference between current time and user's local average rating time.Finally,according to zhou's result about user evaluation of the memory effect which illustrates user's current rating will influence rating score in next peroid of time.With the conclusion above,we district each user's memory regions before training model,and the weight of user's drifting is evaluated by attribute of the memory area.