Research Of A Recommendation System Incorporating Project Support Information

In recent years,with the rapid development of big data and other technologies and the explosive emergence of Internet applications,people are in a society of massive information and cannot quickly and effectively filter out the information they need,while the recommendation system can make personalized recommendations for users by analyzing their historical behavior data.The current recommendation algorithm based on graph neural network fully considers the interaction between users and items and achieves good recommendation results,but due to a large amount of item data,the interaction between users and items has the problem of data sparsity,which leads to the lack of recommendation accuracy;at the same time,users cannot obtain effective preference information from the item collection.To address the above problems,this paper proposes a method to alleviate the data sparsity problem by using item-assisted information enhancement while retaining user-item interaction information for the recommendation,as follows.Firstly,a user-item bipartite graph and an item-item homogeneous graph are constructed based on the historical sequence of user-item interactions,on which the idea of knowledge distillation is introduced to propose a double-tower recommendation model BiInfGCN based on synchronous learning.the model is composed of two modules,the user-item training model and the item-item training model left and right,during the model training process,both sides of the model structure learn from each other and thus improve the final precision on the recommendation task.Secondly,considering that the simple use of item-embedded representation averaging to obtain the user-embedded representation will lead to a certain degree of loss of information about the user’s preference for the item,a recommendation model BiInfGCN-time with a time-decaying distribution function and a double-tower recommendation model Bi Inf GCN-att based on the user’s attention mechanism are proposed.when obtaining the user embedding representation from the item embedding representation,the user embedding representation is obtained through In the case of obtaining the user embedding representation from the item embedding representation,the weight parameters of different items are obtained through the time distribution function or the attention weight matrix,so that the user embedding representation can better reflect the preference level of the items.Finally,relevant experiments are conducted on three real datasets,including Automotive and Last.fm,and the proposed model is compared with the comparison model as well as a comprehensive analysis to verify the validity and reasonableness of this model.