Research On Recommendation Algorithms By Combining Side Information

With the development of the internet, there are more and more kinds of products, the structure of information is becoming more and more complex, as well as the side information that affects the performance of recommend system. The side information includes all other kinds of information such as text, attributes, location, time except the rating matrix. How to introduce the side information into the recommend system is a hotspot of research. One way is to considering all the side information to build one model, but the drawback is that the model is too complex and bad scalability. Referencing to ensemble learning, our research is to ensemble several side information as to improve the performance of recommendation system. The main work of this paper are as follows:First, by modifying CAT (Content+Attribute) algorithm, we trained three single recommenders, one ensembles content, one ensembles attributes, the other ensembles attributes of a user. By comparing to the experimental results of the PMF algorithm, the stability and convergence of these single recommenders are proved.Second, by trying three ensemble algorithms, which are custom weighted method, self-learnt weighted method and Stacking method, this paper ensembles three single recommenders mentioned above, each of which ensembles one kind of side information. Experimental results proved that ensemble results have higher accuracy than that of single recommenders. Comparing to the results of the recommender which has all the side information in one model proved the ensemble algorithms have better scalability.Third, this paper studied user’s evolution pattern of interest, which is a kind of implicit temporal information, and proposed an algorithm which clusters users by their evolution patterns of interests and then does recommendation on each cluster, noted as RUC (Recommendation by User Clustering). By comparing the recommending result with that of the non-clustering recommender, it proved that RUC has higher accuracy and better performance.