Research On A Recommendation System Integrating Neural Networks And Attention Mechanisms

Aiming at the problem of ignoring the time sequence between data and the static user interest learned in the recommendation system of real scene,a sequential recommendation model combining attention mechanism and GRU cyclic neural network was proposed to model the user behavior sequence explicitly and dynamically,and to mine the user’s long-term and short-term preferences.Combined with the user’s own information,an adaptive weighted gated unit is constructed to combine the long-term preference and short-term preference,so as to predict the user’s next possible behavior.The experimental results on Amazon data set showed that compared with the current benchmark recommended models such as GRU4 REC,STAMP,SASREC,etc.,the normalized cumulative loss gain(NDCG)and HIT ratio(HIT)of the proposed model increased by at least 14.7% and 8.8%,indicating that the model can capture user interest more accurately.At the same time,to solve the problem of cold startup in the recommendation system,the article starts from the point of efficient utilization of item information,constructs the article relationship diagram through the graph neural network and graph attention mechanism,so as to excavate the relationship between items and generate the accurate id vector for the newly added item in the recommendation system.According to the experimental results on Amazon data set,compared with the current cold-start strategies such as Meta Emb,Rnd Emb,Ngb Emb,etc.,the accuracy index of AUC of this method is improved by 2.9%,and that of Loss is decreased by 4.1%.It shows that the strategy can generate a suitable id embedding vector for the newly added item,which helps the recommendation system quickly through the cold startup phase.