| The recommendation algorithm is an important method for alleviating information overload in the Internet era.It can screen the content of interest in large amounts of information on the Internet.Classic recommendation algorithms usually learn their personalized interest preferences from the historical clicks of users and predict their possibility of interest in candidate content.However,in the real scene,new users enter the system daily.The model can only obtain a small number of initial clicks from each new user,so it is difficult to generate accurate and reasonable recommendation results in time.This is also known as the problem of user cold-start problem.With the development of the relevant research of the meta learning,it has been widely used in the scenes of lack of samples and achieved significant results.The study of this paper focuses on the application of meta-learning to effectively alleviate the user cold-start,thereby improving the performance of the recommender system.Specifically,this paper first reveals two critical issues of the model faced by applying meta learning in the recommender system and proposes a pre-trained network modulation and task adaptation algorithm PNMTA.Then,this article further studied the challenges encountered in actual applications of meta-learning recommendation algorithms and proposed a deployable and continuable meta-learning recommendation algorithm DCMR.Model-Agnostic Meta-Learning(MAML)has been widely used to alleviate the problem of user cold-start in recent years.Most of the current related work regards generating recommendations for a user as a task.Still,it has led to the unexpected problem of user-side data streams,which significantly affects the effect of representation learning of users and reduces the model’s performance.In addition,these methods usually generate a standard meta model for all users.Still,users with apparent differences in the distribution of preferences will not only damage the knowledge learned by the meta model but also make it difficult for the meta model to adapt accurately to new users.Therefore,to solve the above two problems,this paper designs a pre-trained network modulation and task adaptation approach(PNMTA)to solve the cold-start problems in the recommender system.Regarding the random problem of the user data stream,PNMTA introduced an additional embedding module to modulate the embedding layer parameters that are trained on the traditional algorithm.For user differences,based on the principle of performance distribution,PNMTA uses an additional prediction network modulation module to modulate the parameters of the prediction layer and then adapt to the user.Comprehensive experiments based on two benchmark datasets show that PNMTA has achieved significant and consistent improvements compared to other methods.The existing work thoroughly verifies the advantages of MAML in alleviating the cold-start problem.In the actual application,new users continually enter the system over time,so the recommendation system needs to be able to continually update with continuous and stable users to learn the knowledge of new users continuously.At the same time,the system should also have high-efficiency and good applicability.For this series of issues,this paper proposes a deployable and continuous meta-learning recommender system DCMR.Specifically,DCMR is based on the incremental learning framework of sampling and replay.At the same time,according to the characteristics of the meta-learning recommender system,this paper has designed a dual-constrained task sampler,distillation-based loss functions,the parameter group update rules,the first-order incremental meta learning method,dynamic update controller,and other modules.The experiments performed on the three benchmark datasets fully illustrate the superiority of DCMR. |
