| The session-based recommendation is a rising research field in recommender systems.The session-based recommender system divides the user's sequential interactions into sessions,and recommends the products that the user may be interested in the next step according to the interest preference the user is showing in the current session.The session-based recommender system based on a recurrent neural network is one of the most popular architectures now.The main idea of the system is trying to utilize the excellent serialization modeling capabilities of the recurrent neural network to capture the user's short-term preferences in the current session,and thereby offering the recommendation suggestions to the user.However,this method only focuses on the user's short-term preferences but ignores the user's long-term habits in the past sessions,that leads to the cold start problem which often happens in the session-based recommendation based on a recurrent neural network.Although some scholars had proposed some solutions to improve the cold start problem,however,those methods only use another simple recurrent neural network to model the user's long-term historical interactions,which means they cannot capture the difference in the importance of different sessions and products to the user.In order to solve the shortcomings of the above-mentioned session-based recommendation models,this paper proposes a new session-based recommendation model based on a hierarchical attention mechanism.The model adopts a two-level network structure,the first level of the network is responsible for modeling the short-term interactions of the user in the current short session,it uses an extra user embedding vector as an input to the recurrent neural network,thus the network can capture more personalized short-term preferences of each user,thereby improving the accuracy of recommendations.The second level of the model is responsible for modeling the long-term behavior of the user in past sessions,it consists of two recurrent neural networks with attention mechanism,one is for capturing the user's main purpose in each past short session,and the other one is trying to figure out the different importance of the user's past sessions.Thus,the model can extract a more accurate and personalized long-term behavior feature vector for each user,and everytime a new session starts,the model would use this feature vector to initialize the hidden state of the first level network.In this way,the user's long-term interaction habits and short-term preferences are better integrated to improve the cold start problem in traditional session-based recommender systems,and hence improve the performance of the model when they're recommending items to the user.This paper carried out extensive experiments on two public datasets provided by Reddit and Last.fm.The results show that the proposed model outperforms the state-of-the-art baselines in terms of Recall@20 evaluation metric on both datasets.It achieves a 1.42% improvement over the best performing comparison model on the Reddit dataset,and holds a 1.94% improvement on the Last.fm dataset.In addition,the results show that on both datasets,compared with the existing cold start solutions,the average Recall@20 of the proposed model in the initial stage of each session is increased by about 3%. |
