A Research On Anti-migration Method For Reinforcement Recommendation

As Internet companies need to create a complete ecosystem of products to form a closed-loop flow,more and more platforms develop their own businesses horizontally such as "video + e-commerce","music + social" and other multi-domain platforms.However,the recommendation system in the new field faces the problems of fewer samples and long timeconsuming retraining and deployment.On the other hand,the recommendation model in the new field requires continuous learning and evolution.Therefore,it is necessary to establish a recommendation mechanism that quickly uses the existing mature domain recommendation model to recommend new domains and continuously learns new domain tasks and user interests.In response to the above problems,a set of recommendation models combining transfer learning and reinforcement learning are designed.We use transfer learning to achieve crossdomain initialization of mature models and reinforcement learning to achieve continuous evolution of models.Migration learning designs a network structure composed of encoder,decoder,and discriminator.It finds the shared feature space as much as possible while limiting the degree of feature conversion.As a result,the knowledge of the source domain is retained as much as possible and the common characteristics are fully explored.Aiming at the negative transfer problem in transfer learning where some source domain knowledge is not helpful to the target domain model,a discriminative gating mechanism is introduced to control the transfer ratio.The overall framework of the combination of transfer and reinforcement is proposed to realize the initialization of reinforcement learning model.The reinforcement learning is used to continuously improve the recommendation quality in the target field following the user's interet.According to the needs of the framework,the vectorization and clustering methods are determined,and the specific content of the three elements of reinforcement learning action,reward and status are designed.Based on the above model,a multi-domain-oriented recommendation system is designed and implemented.The system can provide users with video and e-commerce coexisting information services and recommendations of related content.At the same time,it provides a visual display of algorithms for researchers.We use the crawled Youtube and Amazon data sets to test the algorithm.Compared with mainstream domain confusion algorithms,it can have higher recommendation accuracy in terms of performance indicators.Hit Rate increased by about 5.6%,NDCG increased by about 9.4%,and F1 increased about1.06%.The algorithm efficiency can be improved by about 16.2% compared with not introducing transfer learning.