Research On Session Recommendation Method Based On Deep Learning

In recent years,big data,Internet of Things,mobile Internet and other technologies develop quickly,and data scale increases greatly.In the background of big data,information overload becomes normal behavior.Under this background,users and information providers face up many problems: For information providers,it is a difficult problem to show their information to users in a reasonably way depend on users' needs to improve their customer experience and satisfaction.For users,it is also a very difficult problem to find the information that they are really interested in a quick and accurate way from the big data.In order to solve above problems in the background of big data,recommendation system comes into being.Recommendation system aims to help users to get the information that they are really interested in quickly and accurately from the big data.In the traditional recommendation system,it works depended on users' personal information and users' historical records.However,as a matter of fact,many details are difficult to obtain,because it is difficult to get users' personal privacy information.In this situation,session recommendation is proposed.Due to users' privacy information protection,users' detailed information and historical transaction records cannot be obtained in the e-commerce platform scenario.However,session recommendation only needs users' current session information(the product sequences that users click under the current e-commerce platform)to generate product recommendation lists for them,which is very suitable for the scenario.Among the old session recommendation methods,which based on deep learning to have achieved successful results.However,among the old session recommendation methods,there are still the following deficiencies: First,the old session recommendation methods ignore the similarity of different user interests when recommending,and add irrelevant products to users,which increase the amount of calculation and also affect the accuracy of the recommendation results.Second,the old session recommendation methods ignore the relationship between the length of time and the users' interest in the product.The third is that in the old session recommendation methods,the relationship between the indirectly connected commodity nodes in the session is ignored.In order to solve above problems,this thesis studied from the following two parts.Firstly,an adaptive weighted rough K-means was proposed.The method could cluster sessions depend on similar interests more accurately.In the process,recommendation based on session clustering could greatly narrow the scope of the recommendation and improve the accuracy of the recommendation.Second,on the basis of clustering sessions,session recommendation method based on deep learning was proposed.Firstly,for the clustered user sessions,a method to strengthen the users' interest orientation in the sessions was proposed and aimed to strengthen their interest.After that,the products in the sessions were constructed into a product network.Then this thesis used the graph neural network to obtain the feature vectors representation of the commodity nodes information.In the process of getting the feature vectors,not only considered the influence of the node's direct neighbors on the nodes,but also the influence of indirect neighbor nodes on these nodes.Finally,the feature vectors of the products were used to represent the interest characteristics of the users,and then the probability of users' clicking the candidate products were calculated,and users' recommendation lists were generated according to the probability.This thesis compared the experiments with the classic session recommendation methods on the indicators P @ 20 and MRR @ 20.The experimental results showed that the method proposed in this thesis had better recommendation effect in session recommendation.