Research On Recommendation System Based On Graph Neural Network

In the context of the era of information explosion,recommendation systems have emerged to solve the problem of information overload and provide personalized recommendations for users.However,there are a large number of non-Euclidean data in massive data,such as graph-type data.Traditional sequential recommendation algorithms are difficult to obtain spatial information from graph-type data,while sequential recommendation algorithms based on graph neural networks are widely used in academia and industry because they can better capture the high-order interaction relationships implied in graph-type data.However,existing graph neural network sequential recommendation algorithms still have some shortcomings: lack of processing for sparse sequential data,ignoring the interest features of user and item dynamic changes.Therefore,this thesis uses heterogeneous graph neural network model and dynamic graph neural network model to solve the above problems.The main work includes:1.This thesis proposes a heterogeneous graph neural network sequential recommendation model that combines local interest information and user interest information of sequences.The model includes:(1)Constructing a heterogeneous graph with two types of nodes and three types of edges,obtaining node representations after propagation of different types of edges,solving the problem of insufficient interaction information in sequential data.(2)Using an attention mechanism based on input items to combine position information in sequences,obtaining the local interest representation of the sequences,then using an attention mechanism to combine sequence item representations and user long-term representations,obtaining user interest representations of sequences,and finally using a gate mechanism to obtain final representations of user sequences,combining candidate set sequences to complete recommendations,achieving the purpose of improving candidate set sequence recommendation performance.2.This thesis proposes a dynamic graph neural network sequential recommendation model based on multi-head attention mechanism.The model includes(1)construction of dynamic graphs and subgraph sampling methods for dynamic graphs,obtaining graphs with temporal information(2)using multi-head scaled attention mechanism,combining temporal vectors with user and item node representations,obtaining long-term interest representations,then using attention mechanism to combine user item node representations with user item sequence last item representation,obtaining short-term interest representation,finally integrating long-term and short-term interests to achieve personalized recommendation for candidate set sequences.On MovieLens-1M dataset,Amazon’s Beauty dataset and CDs dataset,comparative experiments and ablation experiments were conducted on the two models proposed in this thesis,verifying the recommendation performance of the models and analyzing the impact of hyperparameters on model performance.Finally,this thesis designs a movie recommendation system based on the core algorithm of graph neural network sequential recommendation model.The system mainly includes movie details,popular recommendations,guess your favorite and other display functions.