Research And Application Of Cross-domain Recommendation Algorithm Based On Transfer Learning

In the age of information explosion,people are easily overwhelmed by the amount of data information.Millions of new products,texts and videos are being released all the time,making it hard for people to find anything they're interested in.Users' historical interactions are thought to contain rich information about their interests,which can be used to predict their future interests.This has led to the emergence of recommendation systems.Although the recommendation system has been proved to play an important role in various applications,there is a problem that greatly limits the performance of the recommendation system.The number of user interaction records is usually small,which is not enough to mine users' interests well.This is the problem of data sparsity.As more and more users begin to interact with multiple domains,such as music and books,the possibility of using information gathered from other domains to mitigate this problem increases.This is to use the idea of transfer learning for knowledge transfer to achieve cross-domain recommendation.Compared with traditional recommendation system,cross-domain recommendation is more complex.Although people have made a lot of efforts in the field of cross-domain recommendation,they still face the following problems: the internal relationship between auxiliary data and target data is not fully explored,the information transfer is insufficient,and the recommendation performance is poor;The problem of knowledge transfer is mainly solved by aligning the distribution of data in source domain and target domain.These methods require access to both source and target datasets,however source domain data may not be available in certain scenarios due to privacy regulations or transport restrictions.This thesis proposes a cross-domain recommendation algorithm TCDR-MIB based on information bottleneck,which combines the idea of information bottleneck with transfer learning,and relieves the problem of data sparsity in recommendation system.In this method,the domain-invariant features and domain-specific features are learned separately to reduce the difficulty of model learning.Firstly,the domain-invariant features of overlapping users in source domain and target domain are extracted by using information bottleneck principle,and then item content information is introduced as domain-specific features to assist learning.In addition,for the problem that source domain data is not visible in specific scenarios,the thesis proposes a collaborative denoising cross-domain recommendation algorithm CDCDR considering source-free scenarios.This method fully considers the noise problem of implicit feedback data and combines the characteristics of noise data in the process of deep learning.In the training process of target domain model,the source domain model is used to filter noise data for target domain.In the same strategy,the target domain model is used to filter noise data in the fine-tuning process of the source domain model,so that the source domain model and the target domain model can carry out collaborative training,and realize knowledge transfer to improve the model recommendation quality.Experiments are carried out on several common data sets of recommendation tasks,and the feasibility and effectiveness of the proposed algorithm are verified by comparing existing algorithms.