Research On GAT-based Matrix Completion And Regularized Social Recommendation Algorithm

The embeddings learned by most existing matrix completion methods cannot generalize to new matrices.In order to make matrix completion inductive,most previous works use side information to make predictions.However,high-quality side information is not always available.It is a huge challenge to establish an inductive matrix completion model without relying on side information.In the face of data sparsity and cold start scenarios,the performance of the matrix completion algorithm will also drop sharply.In addition,building social recommendation based on GNN faces the following challenges:using user-item graphs to encode interactions and their associated opinions;strong and weak ties are mixed together in social networks;users involve two graphs(interaction graph and social graph).As a novel convolutional neural network,the Graph Attention Network(GAT)uses the attention mechanism to implicitly assign different weights to different nodes in the neighborhood,and does not rely on knowing the entire graph in advance.The structure retains a very complete locality and is applicable to inductive learning.In addition,with the help of matrix factorization technology,GAT can be used to map high-dimensional sparse user/item vectors into low-dimensional dense embeddings.In view of this,this thesis uses matrix factorization as the basis and GAT as the neural network framework to study two types of recommendation algorithms.The main tasks are as follows:First of all,for the problem of data sparsity,we combine the graph attention mechanism with the multi-relation graph convolution operator to learn an inductive matrix completion model(R-GATMC).This not only captures the rich graph patterns introduced by different edge types,but also distinguishes the importance of different neighbor nodes to the central node,and then aggregates more information from the neighbors.Different weight parameters are introduced for neighbors of each relationship.The learned attention weights are separately aggregated for neighbors in the same relationship,and then aggregated again.At the same time,using basis decomposition technology and adjusting the embedding dimension can effectively reduce the number of learning parameters and prevent overfitting.Secondly,in order to solve the above three challenges at the same time,we use three attention networks to build GAT-So R algorithm and learn the user's embedding in the item space,the user's embedding in the social network space,and the item's embedding in the interactive user space.Correspondingly,it differentiates the different interaction effects of different items on the target users,the strength of social relationships,and the different opinions of different users on the target item.And based on the similarity between the user embeddings,the social user information is appended to the social regularization term,making full use of the learned user embedding information to prevent overfitting.Using GAT as a network layer in social recommendations will also bring more interpretability to recommendations.The calculation of the graph attention model retains a very complete locality,so the two algorithms in this thesis can be used for inductive learning.In addition,the graph attention layer does not require expensive matrix calculations and can be parallelized among all nodes in the graph,so it is computationally efficient.Finally,we conducted a contrastive experiment on the two algorithms,analyzed the performance of R-GATMC in dealing with sparse scenarios,and did ablation research on GAT-So R.The experiment proves the effectiveness of the attention mechanism,and shows that the two algorithms in this thesis achieves highly competitive performance with other algorithms.