Multi-aspect User Preferences Based Recommendation Technology Research

With the continuous development of network and intelligent devices, the information on the Internet is becoming more and more rich and various, people can get a variety of information and services on the Internet. At the same time, people gradually find that it is more and more difficult to find the things and services that accord with their personalized demands. Early search technology provides this service for people, but because o f its insufficient "personality", the personalized recommendation technology begins to be discovered and used in various fields, recommendation system is also trying to use a variety of algorithms for mining user's personalized needs.Personalized recommendation technology mainly attempts to predict the user 's rating of the commodities through the history of user 's behavior, and recommends the corresponding goods to the user according to the predicted scores. Current collaborative filtering algorithm based on matrix decomposition shows good performance on the rating prediction, but it cannot use other auxiliary information(such as user's personal interest and the social relations, etc.) very well, also because of its low dimensional matrix cannot be interpreted after decomposition, it is hard to explain the results of the recommendation. Therefore we deeply analyzed the user behavior, and excavated three different dimensions of user's individual preferences, including the user's emotional preference excavated from the user's comment text through sentiment analysis technology, user's interest preference excavated from the user's text comment using LDA topic modeling, and user 's social preferences through the user's social relations. Then these three different user preferences are fused into the basic matrix decomposition model and the model is solved by the stochastic gradient descent method,on the one hand it can improve the accuracy of recommendation results, on the other hand, we can provide users with explanatory recommended results. Finally we designed detail experiments for the two multi-aspect user preferences based recommender model and the experimental results show that compared with the traditional matrix decomposition model, the accuracy has improved, also we can provide some explanations to the recommendation results.