Research On Session-based Recommendation System With Attention Mechanism And Adversarial Training

With the increasing amount of information,recommendation systems play a significant role in information retrieval applications in the internet age.Traditional recommendation systems used to provide personal recommendation based on the historical behaviors of users.However,due to the loss of data and privacy protection,the historical behaviors of users are difficult to gather on time,while the user's current session data is available in real time.Therefore,there is great research significance in session-based recommendation systems.Current session-based recommendation algorithms transform the task into a problem of sequence prediction.It also models the click sequence by recurrent neural network in order to predict the next item users might click.However,the limitation of these methods is that it only focuses on the current click order and the relevance of the last item in this order.The potential correlations among items in the session sequence has been ignored.It would be hard to exploit user interest.Moreover,recommendation algorithms based on deep learning are prone to overfitting in practice which causes poor generalization performance and affects the effectiveness of recommendation.In order to solve the problems above,this paper proposes session-based recommendations system with attention mechanism and adversarial training,SR-AMAT for brevity.The model uses recurrent neural network to encode items,and then uses the self-attention mechanism to assign weights to the items based on the correlations among the items in the sequence,so as to exploit the user's interest and improve performance of the model.Furthermore,the model is set adversarial perturbations in embedding layer based on fast gradient method with initialization.The adversarial training can be utilized as a regularization method to prevent overfitting so that it might further improve the accuracy of recommendation.To verify the effectiveness of the model,the essay performed extensive comparison experiments with baseline models on three real industrial datasets,YOOCHOOSE,DIGINETICA and TRIVAGO.The results of experiments show that SR-AMAT proposed in this paper outperforms the state-of-the-art baselines in terms of MRR@20 and P@20.