User-intention Guided Recurrent Neural Networks For Sequential Recommendation

Recommender System plays a crucial role in the development of e-commerce and resource providing websites.It is also an important way for websites to increase user stickiness and improve their profitability.Recommendation algorithm is the core part of the recommendation system.And the improvement of the recommendation performance and the interpretability of the recommendation results have always been the two hottest research topics in this field.Sequential Recommendation,which analyzes user-item interactions from historical trajectory perspective and model user history data in a sequence manner,is a key subfield of Recommender systems.In sequential Recommendation,Markov Chain and Recurrent Neural Networks are most widely used methods to process history features.After that,algorithms such as Collaborative Filtering or Factorization Machines will be introduced to measure relations between target items and users to predict user's future behavior.In recent years,a lot of work has been focused on the combination of Recurrent Neural Networks and attention mechanisms to enhance expressiveness of the model.In addition,some other works have been done to model sequential recommendation problems with Convolutional Neural Networks or pure self-attention method.These works have greatly improved the recommendation performance,but there are still weaknesses in the ability to express the user-item interaction behavior.The expression of user's sequence behavior pattern is the most important part in the research of sequence recommendation.According to the difference of user behavior pattern,we can divide the sequence pattern into union-level sequence and individuallevel sequence.Along with these two sequence patterns,skip sequence behavior often occurs.For these two sequence patterns,mainstream recommendation algorithms such as Markov Chain and Recurrent Neural Networks both have some weakness in the expression of these sequence patterns.In this paper,Recurrent Neural Networks is used to model the union-level sequence pattern.For individual-level sequence and skip sequence problem,the attention mechanism,which has strong ability to emphasis and ignore,is introduced to the Recurrent Neural Networks.On the basis of combining the two methods,we propose a user intention guided method to enhance the expression ability of the model and enhance the ability of the Recurrent Neural Networks for personalized recommendation.In general,the main work of this paper are as follows:1.On the basis of combining Recurrent Neural Networks and attention mechanism,we propose a novel UIGR(User Intention Guided Recurrent Neural Network),which integrates user's main purpose information into attention network.In this way,the RNNs reap the ability to express the skip sequence behavior of users,so as to capture the different intentions of different users more accurately.2.We combined UIGR with two classical recommendation algorithms MF and Translation-based manner to construct the model UIGR-MF and UIGR-Trans to deal with the sequence recommendation task.The two model both achieved excellent performance results.3.We conducted extensive experiments on two real datasets,Amazon Video Game and Amazon Office Product,to test model proposed in this paper.In the experiment,we firstly compared our model UIGR-MF and UIGR-Trans with several state-of-theart models on two evaluation metrics NDCG@K and HR@K.Compared with other models,our model UIGR-MF got improvement of 11.76%,7.99%(for NDCG@10 and HR@10)and 14.60%,6.10%(for NDCG@10 and HR@10)in the two datasets.