Streaming Recommendation Algorithm With User Interest Drift Analysis

Recommender system is an effective method to solve the problem of information overload.Outstanding achievement has been achieved after the research and application of academia and industry for many years.While the amount of data increases geometrically with the development of mobile internet,the scalability and data sparsity have been two severe problems for conventional recommendation algorithms.Meanwhile,users' preference may drift as the time goes by,a successful recommender system should be able to capture those changes and adapt to them.To face the challenges above,a streaming recommendation algorithm with user interest drift is proposed in this thesis,in which techniques about matrix prefilling,streaming recommender system and forgetting mechanisms are analyzed and improved.(1)A matrix prefilling method with popularity penalty.Conventional collaborative filtering recommendation algorithm may lead to poor recommendation with extremely sparse data.Nevertheless,matrix prefilling method can alleviate the data sparsity problem by prefilling the missing entries of the rating matrix according to the priori knowledge.After analyzing the experimental results of Enhanced SVD(ESVD),a matrix prefilling method with popularity penalty,named ESVD-P,is proposed in this thesis,which based on the hypothesis that the average error of prefilling ratings has a positive correlation with the activity of users and the popularity of items.(2)A genre-based streaming recommendation algorithm.In order to adapt to changes outside,an online learning process is necessary for a recommender system.In the online learning process,model parameters will be updated incrementally according to the coming data.The fitting error problem occurring in the vector-based incremental updating mechanisms is pointed out in this thesis.To avoid this problem,a streaming recommendation algorithm named streamGBMF is proposed in this thesis,which builds the item vectors according to the genre information.In the online learning process,to keep the overall fitting error from increasing continuously,only user feature vector is updated in real time.(3)Two improved forgetting mechanisms.Due to the drifting of users' preferences over time,partial history records can no longer reflect users' latest habits correctly.In order to exclude the influence of these “outdated” data,recommendation algorithms are equipped with forgetting mechanisms.But the valuable information in the history data will be losing,because existing forgetting mechanisms cannot distinguish the long-term preferences from the short-term preferences.Based on the considerations above,two kinds of novel forgetting mechanisms are embedded according to the characteristic of streamGBMF,and these forgetting mechanisms can effectively preserve users' long-term preferences.To evaluate the performance of our proposed model,the experiments are designed on the popular dataset MovieLens 1M,and different algorithms are compared in streaming environment.The results show that our approach effectively improves the recommendation performance by the proposed matrix prefilling method,streaming recommendation algorithm,and forgetting mechanisms.