Personalized Recommendation Algorithm Based on Graph Embedding and Temporal Context Representation Learning

With the rapid development of deep learning technology and the wide application of intelligent recommendation systems, how to apply deep learning technology to intelligent recommendation systems to continuously improve the recommendation effect has always been a hot research direction. User historical behavior information plays an increasingly important role in the evolution and iteration of the recommender system model. Focusing on how to make full use of the time context of user behavior, how to better deal with the sparsity of massive data and system cold-start problem, the personalized recommendation algorithm is proposed. This algorithm represents the time sequence behavior data with the graph network structure, uses the graph representation learning model node2vec to obtain the graph embedding representation of users and items in the hyperparameter space, and introduces IF_ISF for the negative sampling to improve the model efficiency, and then generates the embedding representation of the user temporal context by the sum-pooling method, and also generates the embedding representation of the static attribute information of users and items by the deep semantic matching method based on DSSM, and finally concatenates the feature vector and uses the machine learning model XGBoost for recommendation ranking prediction. The ablation experiments through the comparison of multiple models on the MovieLens_1M public dataset show that the proposed algorithm is better than the baseline algorithm on the three indicators of MAE, Macro_F1 and Micro_F1, which has different degrees of improvement and can effectively improve the recommendation accuracy and satisfaction.