Research On Graph Convolutional Neural Network Recommendation Algorithm Based On Knowledge Graph

At present,more and more methods focus on using graph neural network to mine semantic information of knowledge graph.Mining semantic information of knowledge graph based on graph neural network(GNN)is the current mainstream method.The idea of GNN mining semantic information of knowledge graph is to map the node ID and edge ID of knowledge graph to a one-dimensional space,and fuse the neighbor node embedding and the center node embedding through the graph convolution operation.The attention mechanism is introduced in the graph convolution operation to give the neighbor nodes a certain weight.Most methods neither aggregate the semantic information of neighbor nodes based on the user's specific interests and preferences,nor model the node embedding representation obtained by the knowledge graph learning,and since the input of the mainstream model is only the ID information of nodes and edges,The input semantic information of the model is very scarce.As a result,the user and item embeddings are not fully mined to represent the semantic information of themselves and each other,resulting in poor recommendation performance.In order to solve the first two problems mentioned above,this thesis first proposes a method based on graph attention convolutional neural network(A-GCN),introduces a specific attention mechanism to mine semantic information of knowledge graph,and iteratively aggregates the graph along the path on the graph.The neighbor node information within two hops of each node is modeled based on the embedded representation of the user and the item learned from the A-GCN,and the user-specific feature embedding representation of the item is generated.In this thesis,click-through rate prediction experiments are conducted on Movie Lens-1M and Last.FM datasets.Both AUC and F1 evaluation indicators exceed the current mainstream methods such as Ripple Net,KGCN,and KGNN-LS.Experimental results demonstrate the feasibility of graph attention convolutional neural networks and re-encoding user embedding representations.This paper also proposes a novel recommendation system framework based on neighbor sampling and global iterative aggregation(NS-FGCN)to further overcome the shortcomings of A-GCN and enhance the recommendation performance of the model.Although the recommendation system based on attention convolutional neural network(A-GCN)introduces a specific attention mechanism to capture the user's specific interest preferences,and models the node embedding to generate the user's specific feature embedding representation of the item,it still has three points limit.First,when sampling neighbor nodes,A-GCN adopts random sampling,which gives all neighbor nodes the same sampling probability,but the importance of different neighbor nodes to the central node is not the same,the smaller the degree is.Item nodes with larger degrees tend to be more fully trained and have richer semantic information.Second,when A-GCN aggregates neighbor node information,it only trains the node embeddings involved in the model training set,and one batch of training does not train all nodes in the knowledge graph,resulting in poor model training results.Third,A-GCN is actually a collaborative filtering recommendation algorithm.The input of the collaborative filtering algorithm is only the node ID,and the semantic information of the node ID is very scarce,and the goal of the recommendation system is based on the attribute information of the item and the user's interest.Like to recommend items to users,the recommendation model cannot ignore the background information and attribute information of users and items themselves.In order to solve the above three limitations,NS-FGCN does not use a uniformly distributed probability distribution to sample neighbor nodes when aggregating neighbor node information,but gives neighbor nodes a certain sampling probability according to the degree of the node.NS-FGCN trains the embeddings of all nodes in the knowledge graph in one batch of training,instead of performing graph convolution operations on only part of the graph nodes based on the nodes involved in the training set.NS-FGCN integrates graph convolutional neural network algorithm and Meta R,Trans E knowledge graph distance transformation mining algorithm.While graph neural network mines graph semantic information,it also pays attention to mining the information of graph nodes themselves.We carried out experiments on three public datasets(Movie Len-20 M,Yelp2018 and Amazon-book).The experimental results show that the algorithm in this thesis is superior to state-of-the-art recommendation algorithms such as KGCN and Light GCN,which proves that the algorithm proposed in this thesis has certain advantages.