Research On Sequence Data Recommendation Algorithm Based On GRU

As the problem of information overload becomes more and more serious,the research of recommendation system has become the focus in the computer field.The traditional recommendation system pays too much attention to the classify and rating of a single item,but ignores the essence of the recommendation system:providing users with a product list based on user interest.In response to the shortcomings of traditional recommendation systems,based on the user session recommendation algorithm,a ranking recommendation algorithm that combines XGBoost and Gated Recurrent Unit(GRU)was proposed.This algorithm makes full use of user's historical behavior record,analyzes user's implicit feedback data to mine user's potential interest,so that user experience and the performance of the recommendation system can be improved.The specific research content is as follows:(1)Traditional boosting tree model only pays attention to the training loss,which may leads to a complicated and not feature-extracting conducive data model.In response to this problem,this thesis innovatively introduced XGBoost.The advantage of XGBoost is that the loss function not only defines the data training loss,but also defines the regular items that control the complexity of the data model.At the same time,in order to better model the relevance of commodity items,the Pearson correlation coefficient is introduced,and a GRU serial recommendation algorithm fused with XGBoost is constructed.The experimental results show that using XGBoost for feature extraction has a better MRR index than using tree lifting model for feature extraction.(2)The traditional Dropout network reduces noise data by randomly deleting clicks in the session.Because the deletion behavior is too random,there is a possibility of deleting important clicks.To solve this problem,the traditional Dropout network is improved.The improved Dropout network introduces the concepts of time point T and residence time(Dewll Time).The conclusion is obtained by analyzing the user's click behavior:there is no distribution of user interest at T=2 seconds.Therefore,the improved Dropout network will delete clicks with a dwell time equal to two seconds.The experimental results show that with the improved Dropout network for data processing,the average recall rate is significantly improved compared to the traditional Dropout network,and the data distribution is more uniform.(3)In order to mine more user sequence information,a graph embedding method is introduced.This article uses the relationship between the number of click steps and user interest to innovate and improve the negative sampling standard in the traditional embedding method.Combined with the improvement and innovation of the Dropout network and the negative sampling standard,a GRU ranking recommendation algorithm based on user sessions is constructed.Use the Learning to Rank ranking loss function,combined with the Pairwise method,to form a recommendation list that is positively related to the query document.The experimental results show that the proposed model overcomes the shortcomings of the traditional recommendation algorithm that pays too much attention to the category and score of a single item,and has improved on both Rec@25 and MRR evaluation indicators.