Research On Cross Domain Recommendation Based On Heterogeneous Information Transformation

With the rapid development of online social networks,the amount of online user information data has exploded.How to extract useful information and provide users with personalized recommendations has become a research hotspot.At present,many personalized recommendation methods have emerged.However,traditional single-domain recommendation methods are usually affected by information sparsity problems and cold start problems,and their ability to solve sparsity and cold start problems is limited.Because Internet users usually come into contact with many different types of social networks,such as academic networks,online dating networks,and multimedia networks,research on cross-domain recommendation algorithms has emerged.Cross-domain recommendation can solve the problem of information sparsity and cold start of recommendation that traditional methods usually face,and has attracted a lot of attention in recent years.In addition to using information from the target network,the cross-domain recommendation algorithm can also transfer auxiliary information from the external networks to the target network to solve the recommendation problem,thereby improving the recommendation performance.In different network applications,there are usually completely or partially overlapping users and items.By making full use of overlapping users,the behavior patterns of the same entity users in the two platforms can be explored,thereby enabling the information migration.However,cross domain recommendation algorithms also face some problems,such as the heterogeneity of information caused by different social platforms.Information heterogeneity is mainly caused by domain-specific differences,that is,in addition to users' personal preferences,due to the different functions of different domains,users' behavior patterns and preferences also differs;at the same time,there are different granularity of information contained in different platforms,such as the level of classification of tags,the length of text information,and so on.In view of the above problems,this thesis proposes a cross domain recommendation algorithm HITCDR to improved cross domain recommendation result from two aspects.First,an inter-domain user alignment algorithm RPSUA is proposed to optimize the cross domain mapping function training,;then an information migration method that distinguishes between domain shared features and domain specific features is proposed.The main tasks completed in this paper are as follows:(1)This paper first generates the user representation in a single domain through matrix decomposition,and combines the user-item interaction matrix to calculate the similarity between users,and builds the user similarity relationship network according to the similarity;then project the user representations from both source domain and target domain to the same representation space;finally,the user pairs with similar representation resultsare marked as implicitly overlapped users,and combined with the identified explicit overlapping users for the subsequent feature differentiation stage and cross-domain representation mapping stage to enhance the performance of the cross-domain mapping function;(2)Then this paper extracts the domain shared features and domain specific features in the user representation through the domain separation network,and uses this to map the two different types of features with different granularity.After training the feature mapping network,the single domain representation of the non-overlapping users in the source domain is used as input and the target domain representation of the non-overlapping users is generated;finally,the newly generated user representation is used to restore the predicted rating matrix.This thesis first introduces and classifies the existing cross domain recommendation algorithms,which lays the foundation for the design of the cross-domain recommendation algorithm in this thesis.Then the design and implementation of the HITCDR in this paper are explained in detail,and the validity of the scheme in the real Amazon data set is verified and an evaluation analysis is given.