Research On User Dynamic Interest Model In Recommender System

The exponential increase of information makes it a hot research topic to recommend highquality content for users with intelligent algorithms and user characteristics.However,it poses challenges for the design of recommendation algorithms,due to the frequent changes of user interest.Non-time-sensitive algorithms cannot instantaneously capture the prediction deviation caused by user's dynamic interest in the actual recommendation scenario.In addition,most time-sensitive methods learn the dynamic behaviors of users,mainly by complex artificial features and multi-models ensembling,which imposes a burden on the design of the model and its online updates.Thus,the design of end-to-end and time-sensitive recommender algorithms is of great research value,with regard to optimizing the accuracy and simplifying the design of the recommended algorithms.The thesis analyzes the dynamics of the recommendation system qualitatively and quantitatively,and proposes that the user's dynamic interest is impacted by user's long-and short-term interest.Besides,there are also user's accidental interests,which may cause interferences to the prediction of user's dynamic interest.Based on the theory,we propose an algorithm to predict users' rating behaviors.In the proposed method,a non-time-sensitive algorithm is used to analyze the user's static interest performance with his/her historical behaviors,in order to obtain the user's benchmark ratings.Then based on the recurrent neural network,the user's dynamic interest is explored to calculate the deviation of the prediction of ratings to amend the benchmark ratings.The main contributions of this thesis are as follow: 1)We propose Long-and Short-Term Interest Networks(LSIN),consisting of a short-term interest network and a long-term interest network.The short-term interest network,based on deep recurrent network,extracts the shortterm interest features,while the long-term interest network is to extract the long-term interest features.And then it fuses the long-and short-term interest features to capture the dynamic interest feature of users,which is used to calculate the current rating deviation.2)We propose a new method to fuse user's long-term interest features and short-term interest features,which combines the long-and short-term interest features by their cosine similarity with the current movie features.Our proposal fusion method achieves higher accuracy while extracting the dynamic interest features.3)Based on the proposal LSIN,we use the attention mechanism to assess the influence of recent behaviors on the current interest,so as to reduce the interference of user's accidental interests,and improve the remote dependence issues of the recurrent network.4)We use multiple datasets split from the Netflix dataset to test the performance of the proposal model.Compared with the current mainstream recommender algorithms,our proposal methods offer better prediction accuracy on all test sets.The RMSE performance of our final model on the Netflix full dataset is 0.9119,which is 2.2% and 1.8% lower than the non-time-sensitive algorithm PMF and the time-sensitive algorithm TimeSVD++ respectively.