Research And Application On Recommendation Algorithm Via Unstructured Information Modeling

We are living in the information era which brings us convenience and helps us obtain the desired information on the internet.The flood of information came in the same time,people can not find what they need efficiently from the huge amount of information.Recommender system learns from user profiles and interactions.It can help people find content that they are interested in.There are many recommendation scenarios which includes sequential recommendation and social recommendation.The correspoding research directions are sequence modeling and relation modeling,which are also important research areas of recommender system,they utilize these unstructured information to make recommendation better.It is of great significance and difficulty to combine scenario property and improve recommendation accuracy.In this paper,we focus on sequence modeling and relation modeling and propose our work.In sequential recommendation,user preference changes over time.The recommendation list and timestamps are both important optimization objectives.Traditional sequential recommendation approaches hardly make good use of timestamp information.Meanwhile,the two objectives may be conflict during optimazation.To improve the above problems,we propose Two-Way Gated Recurrent Unit(TW-GRU)based on sequence modeling.The proposed model dealts with item embedding and time embedding by different optimazation objectives and finds there pareto optima in the form of multi-task learning,finally finds the optimal weights for the two tasks.Another problem is that simply predicting timestamps may not be the best way to describe user preference distribution over time.To make predictions more interpretable,we propose Point-Process Gated Recurrent Unit(PP-GRU)which improves the multitask learning part of TW-GRU by replacing the time prediction task with conditional intensity prediction task.The model can provide recommendation intensity representations in a period of time and make the predictions more interpetable.In social recommandation,there are explicit social connections and interactions,but they hardly exist in traditional rating prediction scenarios.Rich and varied user rating preferences illustrate that users with similar interactions may have very different rating preference.In traditional appraoches,implicit interactions mostly are encoded into a bias term and help the rating prediction.However,they are important to measure the inherent interaction preference.In this case,we propose InteractionDecompose Graph Convolutional Network(ID-GCN)which models implicit interactions into a user interaction graph and uses node embeddings to improve rating prediction performance and immunity.The proposed model leverages between inherent implicit connections and rating performances to make predictions more robust.Finally,we present our experiments on e-commerce log dataset RSC15 and movie rating dataset Movielens.We compare our proposed model with traditional recommendation models to testify the effectiveness and superiority.Moreover,we present additional tailored experiments including multi-task learning analysis,event intensity prediction and immunity comparison for the three proposed models to evaluate influence of their respective improvements.