Research And Implementation Of Session-based Recommendation Algorithm Based On Deep Neural Network And Graph-based Embedding

As an effective technology,the session-based recommendation system can help users find items related to their interests in the current session.Compared with traditional recommendation systems,session-based recommendation is highly practical because it can make recommendations based only on user behaviors observed during the current session,and has received extensive attention from researchers.The current research mainly models the session-based recommendation task as predicting the user's next click problem based on the existing click sequence data,using deep neural networks and representation learning methods to extract information related to user interests,thereby calculating the probability of being clicked on all candidate items.This thesis analyzes and summarizes the existing mainstream related work and finds that there are some problems: First,the user's interest in the session has various and changeable characteristics,resulting in the current mainstream neural network model modeling the user interest insufficiently,which affects the accuracy of recommendation;Second,current work focuses on how to effectively model the user behavior,but ignore the influence of the information contained in the item embedding on the model,and lack the use of the item information and the complex relationship between items,which restricts the further improvement of the model performance.In response to the above problems,this thesis takes the user's anonymous session recommendation as the research object,and proposes two session recommendation algorithms.First,aiming at modeling the human browsing behavior,this thesis proposes a short-term attention and memory priority session-based recommendation algorithm that prioritizes the user's current interest in the session.And an attention mechanism is designed to capture the features related to the user's long-term interest and current interest simultaneously to synthesize the session context representation,which alleviates the problem of user interest drift that affects the accuracy of recommendation,which is difficult for current session-based recommendation algorithms to handle.Meanwhile,summarizing the shortcomings of using randomly initialized item embeddings in the current recommendation algorithm,innovatively constructing an item access frequency and sequence relationship graph based on all session data,and designing a graph-based item representation learning method.The concise and efficient way of semantic synthesis learns the popularity of items at the global level and the rich and complex associated information between items.Finally,proposing a graph embedding based session recommendation algorithm,which pre-learns the fixed vector representation of all items from the graph as the input of the recommendation model,so that it not only has rich information such as the item itself in the global attention situation,but also retains the complex relationship between items at the global level.In addition,when modeling the session context,both the long-term behavior of the user including the general sequence information and the short-term behavior including the user's initial and current interests are combined to improve the overall performance of the recommendation algorithm.A series of experimental results on three real public datasets show that the overall performance of the proposed session-based recommendation algorithms have reached the state-of-the-art level.In addition,experiments have proved that using the learned fixed item representation as the input of the recommendation model can effectively improve the adaptability of the algorithm in real scenarios.And this thesis also provides new research ideas for the field of session-based recommendation.