Research On Recommendation Algorithms For Multi-task Learning With Directed Graph Convolutional Networks

As an important branch of machine learning,recommendation algorithm has attracted the attention of many experts and scholars.Especially in the field of recommendation,the recommendation algorithm plays an important role as a bridge connecting users and products.Whether it is a traditional recommendation algorithm,such as collaborative filtering recommendation algorithm,content-based recommendation algorithm or hybrid model recommendation algorithm,there are more or less problems such as cold start and simplification of recommended items.Therefore,in order to overcome these problems to improve the accuracy of personalized recommendation algorithms,the following research is carried out in this paper.Firstly,this paper proposes a recommendation model for Multi-Task Learning based on Directed Graph Convolutional Network(MTL-DGCNR)based on directed graph convolutional network and applies it to the recommendation field of e-commerce.This model is first constructed The user’s micro-behavior,using the user’s micro-behavior embedding helps to better capture user preferences at a fine-grained level,and then convert the user’s micro-behavior into directed graph structure data for model embedding,and fully consider the first-order adjacent nodes and The embedding of second-order adjacent nodes can effectively enhance the transformation ability of features and obtain a simpler distribution than graph convolutional neural network(GCN).In addition,using the directed graph convolutional network as the main task for recommendation can avoid the problem that the recurrent neural network(RNN)only models the single transfer relationship of two adjacent items in the recommendation,thereby ignoring the problem of other item information in the session.Secondly,the MTL-DGCNR model adopts the multi-task learning method,and uses the learning embedding of knowledge graph to effectively deal with the one-to-many or many-to-many relationship between users and products,so that the user’s conversational representation is clearer,and as a The auxiliary task assists the main task DGCNR model to complete the recommendation task,and this method can effectively alleviate the cold start problem existing in the recommendation task.Finally,through comparative experiments with eight existing hybrid recommendation models on the real e-commerce data set JData,JData(N)with cold start items,and Retailrocket e-commerce website behavior data,the results show that MTL-DGCNR is better than other models.The recommendation model has higher interpretability and accuracy in the field of e-commerce recommendation.In addition,the ranking evaluation experiments,various training methods comparison experiments and control parameter experiments designed in this paper can verify the rationality of MTL-DGCNR from multiple perspectives.