User Behavior Analysis And Personalized Recommendation Based On Graph Neural Network

In the current era of big data,recommendation systems have begun to be used in various scenarios,especially in e-commerce platforms,news information,video websites and other scenarios.For the application scenario of video websites,this paper analyzes the user’s short-term interactive behavior data,user personal feature data and video feature data,and uses two stages of recall and sorting to generate a personalized video recommendation list for each user.To quickly narrow down the candidate set,the sorting stage is used to fine-sort the recalled candidate set.In the recall stage,a network graph with video as a node is constructed based on the user’s interactive behavior data,and the Node2 Vec algorithm in the graph neural network is used to extract the feature representation vector of the video,and then the similarity matrix between videos and videos is obtained.The video interaction matrix is collaboratively filtered,and a video candidate set of size 200 is initially screened for each user.In order to study the change of user interest with time,this paper improves the original interaction matrix by introducing time exponential decay.Through empirical comparative analysis,it is found that the value of recall after adding the time decay factor is 0.297124,which is an increase of 1.4% compared with the unimproved model,indicating that the change of user interest is an important factor in the analysis of user behavior.In practical applications,it is necessary to model the user’s interest according to the length of time.In the sorting stage,based on the user feature data and video feature data,the sorting model is trained with the goal of whether the user watches videos.And the logistic regression,XGBoost,Light GBM and Deep FM models are trained respectively.In this paper,the model fusion method is used to obtain the final ranking model.Compared with the single optimal model,the AUC index of the fusion model is improved by 1.296%.In the actual modeling process,this paper conducts feature engineering on user features and video features,constructs features such as user interest preferences and user activity from the perspective of users,and constructs video popularity from the perspective of video features.Through the feature importance analysis of the XGBoost model,it is found that the user interest preference feature constructed based on the user interest portrait is the most important feature,indicating that the features constructed in this paper are necessary for the prediction of results.Finally,this paper uses the model trained in the sorting stage to sort the video candidate set obtained in the recall stage,taking the first 30 videos as the user’s Top-30 personalized recommendation list.Obtain the combination of the recall model and the sorting model after adding the time decay factor reached the highest recall 0.235789.