The Research And Implementation Of Recommendation System Based On Graph Embedding And Multi-tasking Learning

With the increasing amount of network information,the recommendation system is becoming more and more important as a tool for solving information overload.The core part of the recommendation system is divided into two modules,recall and sort.The recall module needs to quickly retrieve the candidate items which may be of interest to the user from the original set of items.The sort module accurately sorts the recall set for recommendation.This thesis studies the recall module and ranking module in recommendation,and mainly completes the following three aspects of work:(1)The graph neural network GraphSAGE model is introduced into the recall stage,and the users' and items' nodes are represented by graph embeddings.The recall operation is based on the inner product of the embedding vector of the item and the user.In addition to using the aggregation method in the original GraphSAGE model,this thesis also proposes an aggregation method based on the attention mechanism.Recall experiments were performed based on the MovieLens-20M movie and Electronics product datasets.The experiments verified that the hit rate of the proposed recall algorithm is significantly higher than the current mainstream deep model of dual-tower structures and the EGES method based on graph embedding.It is also-verified that the aggregation method proposed in this thesis is better than the aggregation method in the original GraphSAGE model.(2)Multi-task learning is introduced into the recommendation ranking stage.The recommendation task of a single user is modeled as a separate regression or classification task,and the recommendation tasks of all users are trained jointly using a meta-learning algorithm.This modeling method can be combined with many current deep recommendation network architectures to get more accurate recommendations.The experiment is based on the four different specifications of MovieLens and the Electronics product datasets.The comparison is based on several current recommendation network architectures.The results show that the recommendation model using multi-task learning modeling can have better performance.And the degree of performance improvement increases with the number of users in the dataset.(3)A specific personalized movie recommendation system is constructed based on the Flask framework.This system uses the recall and ranking modules described in the first two parts of this article,which can make personalized recommendations based on the user's historical movie viewing records.