Research On Cross Domain Sequential Recommendation Algorithm Based On Deep Learning

In recent years,with the rapid development of computer technology and Internet,information system is not only convenient for people to communicate,but also seriously troubled by the problem of information overload due to the generation of massive information.Therefore,personalized recommendation technology which can filter and screen information for users arises at the historic moment.The main task of the recommender system is to discover the potential needs of users based on modeling their interests according to the user's past behavior records.On the one hand,it can help users find the items they may be interested in;on the other hand,it can also make the relevant items displayed in front of the corresponding user groups.Therefore,recommender system has great application value for both producers and consumers.At present,the recommendation system has been widely concerned by all walks of life and the majority of researchers,and has achieved great success.Most of the current recommender systems focus on modeling user interests according to the history of users in this field,but ignore the essential characteristics of users' behaviors in different domains.Therefore,cross domain recommendation based on multi-domain data modeling user interests has become a research hotspot.In this paper,the cross domain recommendation system based on deep learning is mainly based on user behavior sequence modeling user preferences,namely cross domain sequential recommendation.The cross domain sequential recommendation considers both shared account and cross domain,that is,shared account cross domain sequential recommendation(SCSR).The task of SCSR is mainly faced with two challenges:(1)The history of multiple users in the shared account is mixed and difficult to distinguish.(2)User's behavior is generated in multiple domains,and users' consumption of different items in different domains can feed back users' similar preferences.The specific research contents of this paper are as follows:Firstly,a RNN-based cross-domain recommendation model(RCRM)is proposed.RCRM considers both shared account and cross domain.It assumes that users in shared account have similar preferences,and uses recurrent neural network to express the overall interest of shared account.Aiming at the cross domain problem,a multi-layer cross mapping perceptual network(CMPN)is proposed to realize the account information transmission between different domains,so that the information of multiple domains can improve the recommendation quality of the target domain.Then,a self-attention-based cross-domain recommendation model(SCRM)is proposed.For the case of shared account in SCSR task,SCRM introduces the multi-head self-attention network in The Transformer model to model multiple users in shared account.For the cross domain situation in SCSR task,this paper improves on the previous CMPN,and proposes an improved version of cross domain transmission network,namely ICMPN,to better mine users' interests in different domains,make full use of users' information in different domains,and further improve the recommendation quality of the target domain.Finally,this paper makes a detailed experimental comparison and analysis on HVIDEO dataset and HAMAZON dataset.Compared with other related methods,our method can achieve better recommendation results in MRR and Recall.