Research On Session-based Recommendation Algorithm Based On Deep Learning

In the Internet era,the problem of information overload is coming to the fore.As an effective solution,recommendation systems are increasingly important in e-commerce,ad recommendation,short video,etc.Session-based recommendation is a relatively new branch of recommender systems that differs from other recommendation directions.It protects user privacy and captures short-term dynamic user preferences,allowing the model to follow changes in user preferences and improve the accuracy of recommendations.Currently,the research on session-based recommendation is increasing,and the mainstream and best results are based on Graph Neural Networks(GNNs)for session-based recommendation.Unlike RNNs,GNNs can learn higher-order neighbor information and do not have strict sequential relationships.These advantages improve the effectiveness of the recommendation model.Although GNN has made significant progress,it still has some drawbacks and limitations.We put forward the following research work to solve these problems:(1)There is noise in the session data.The current main solution is to use the attention mechanism to eliminate noise,but it is not effective enough because it is a soft pruning denoising method.Attention-based soft pruning is not thorough enough for noise removal.We propose the DGNN model(Denoising Graph Neural Network).It contains two innovations: the proposed Swin Relu function to achieve the masking of noisy messages and the proposed Flow Valve for constructing the threshold value for masking noise.Combining these two points achieves the operation of adaptive thresholding hard pruning,which makes the noise elimination more complete.(2)In session-based recommendation research,extracting the session as a vector as the user’s preference is necessary.The current mainstream practice is to take the last interaction item in the training data as the user’s short-term interest,calculate the attention score as a QUERY with the items in the session,and the weighted sum of the corresponding items as the long-term interest.The long-and short-term interests are combined through a layer of Fully Connected Network as the user’s preferences.In this approach,the long-term interest obtained is more homogeneous due to the short-term interest as a QUERY,but multiple interests often influence users’ purchase behavior.Besides,unlike other recommendation directions,session-based recommendation does not have access to the complete interaction history of users,which makes the model more susceptible to data sparsity.Based on the above,the DAS-GNN model(Denoising Autoencoder integrated with Self-supervised learning in Graph Neural Network)is proposed.First,it uses a denoising autoencoder to construct QUERY for constructing long-term interests,overcoming the homogeneous problem.Second,to address the data sparsity problem,we combine Self-Supervised Learning(SSL)to build positive and negative samples,which allows the model to exploit unobserved item information in addition to observable user interaction data,alleviating the data sparsity problem.(3)It is a common practice in session-based recommendation to use the last item in the session data as the label.In session data,there may be multiple interests of a user,and the label provides a target for model optimization,but the label can only represent one interest of the user.This leads to the possibility that other interests in a session data may be optimized toward the wrong gradient.We proposed the CGNN model,which uses category information in the dataset to distinguish the different interests of a user,so that the labeling optimizes the items corresponding to the interests without affecting the other interests of the user.In this thesis,these three models are carefully explained and analyzed,and experiments are conducted on several datasets.The experimental results show that they outperform other advanced models and validate their superiority.