Research On Recommendation Algorithm Optimization Based On LightGCN

With the rapid development of modern information society,the density and dimensions of information have gradually reached a point where it is difficult to calculate.In order to mine high-quality information from massive amounts of information to meet users’ needs,recommendation systems have emerged.The main idea of modern recommendation systems is based on collaborative filtering,which calculates the similarity between users or items,and then recommends them to similar people or similar items.This method was applied to early recommendation systems.With the advent of the era of information overload,collaborative filtering revealed drawbacks: The sparse matrix constructed when the amount of data is large reduces the accuracy of recommendation results,causes cold start issues due to a lack of historical data,and wastes computing resources due to real-time computing of user and item similarity so on.Facing these problems,scholars put forward the graph neural network model and solved the problems in the history recommendation system by optimizing this model.The Light GCN model learns the embedded representation of nodes by building a graph model and aggregating messages from neighboring nodes of the graph node.It uses high-level information to mine potential features of embedded vectors,so it performs well in recommendation systems.However,this model also has some areas that can be improved:the prediction layer cannot map potential vectors to prediction scores using matrix inner products,and layer nodes have the same weight when aggregating the main research contents of this article are as follows:(1)In view of the inability of the prediction layer to obtain potential vectors using matrix inner products,that is,to capture nonlinear relationships,this thesis proposes using neural collaborative filtering networks instead of matrix inner products to expand low dimensional matrix relationships to high dimensional ones,better capturing the potential relationships between users and items.This process is validated through comparative experiments with the Light GCN model.(2)Aiming at the problem of the same weight between nodes during message aggregation,which may reduce the expressiveness of embedded layer vectors,This article adopts a self-attention mechanism that can capture global embedded information.By learning weight information between different locations,it enhances the expression of embedded vectors.(3)This article compares and verifies the above optimized model on Yelp 2018,Gowalla,and Amazon Book.The final experimental results show that the improved Light GCN model has better performance by adding self-attention mechanism and replacing matrix inner product with neural collaborative filtering model.