Research On Session-based Social Recommendation System

The recommendation system can capture user’s consumption preferences from their interactive behaviors,and bring considerable commercial profits to the platform through personalized recommendation.However,traditional collaborative filtering-based recommendation algorithms can only model user’s static and stable preferences,resulting in inaccurate recommendation results of the system.With the development of deep neural networks,researchers have proposed session-based recommendation with deep learning method,which regards user’s transaction behavior as time-varying session information,then captures user’s dynamic preferences from it.Although this method improves the recommendation accuracy,it relies too much on user’s interaction information and ignores the influence of friends’ interests on user’s consumption preferences.In order to solve the above problems,this paper studies on session-based social recommendation system,and the specific work is as follows:1.This paper proposes a new session-based social recommendation model with recurrent neural network to simultaneously capture user’s preferences and social influences.Firstly,this paper uses the recurrent neural network to design a method for expressing the correlation of user purchase behavior based on feedback mechanism to learn the dependencies between transaction behaviors in recent sessions and capture user’s consumption preferences.Secondly,a multi-layer recurrent network structure is designed,which enriches the feature information of neighbor nodes and deeply mines the interests of social friends.Finally,a graph attention convolutional neural network is used to capture the social influence between nodes to obtain the user’s final consumption interest.The experimental results demonstrate the effectiveness of the proposed methods,their Recall(recall rate)and NDCG(Normalized Discounted Cumulative Gain)on Douban Movie and Delicious datasets both outperform the state-of-art model DGRec.2.Aiming at the problem that the previous session-based recommendation methods rarely consider the influence of user’s global transaction information and friends’ interests on user’s interaction behavior,this paper proposes a session-based social recommendation algorithm AFGSRec that adaptively integrates global collaborative features.Firstly,a heterogeneous graph neural network is used to model users’ historical transaction information in social networks,and capture the global collaborative features and social influence among friends.Secondly,a graph neural network based on selection mechanism is designed,which effectively filters out the node transition information irrelevant to the current session and captures the user’s current preference more accurately.Finally,an adaptive feature fusion method is proposed to dynamically capture the impact of global collaborative features on user’s recent preferences.Experimental results demonstrate the effectiveness and robustness of AFGSRec,which outperforms the state of art model SERec in both HR(Hit Rate)and MRR(Mean Reciprocal Rank)on Gowalla and Delicious datasets.3.The current recommendation model usually uses the grid search method to determine the numerical value of the learning rate during the training process.This method takes a lot of time and cannot guarantee the convergence effect of the model.If the value of the learning rate is too large,the model will not be able to converge in the later stage of training.If the value of the learning rate is too small,the convergence rate of the model will be very slow,and it is difficult to jump out of the saddle point,which will eventually lead to the inaccurate recommendation results of the model.In order to solve the above problems,this paper uses the dynamic learning rate based on the triangular cycle for model training,which will affect the model performance in the short term,but it will help improve the model training effect in the long term.The experimental results on the Gowalla and Delicious datasets prove the effectiveness of this method.It not only improves the recommendation accuracy of the model,but also improves the model convergence speed.