Research On Cold Start Recommendation Algorithm Based On Meta-contrasive Learning

With the development of the Internet,the problem of information overload is becoming more and more serious.As a recommendation system to help users filter a large amount of information and find suitable content,it is becoming more and more important.But when new users and new content join the system,the cold start problem also arises.Traditional recommendation algorithms usually require a large amount of interaction history for training,but in the cold start scenario,user interaction history data is very sparse,and traditional recommendation algorithms cannot learn accurate vector representations,which will also cause serious overfitting problems due to the lack of sufficient supervision signals.At present,the methods to alleviate the cold start problem are mainly divided into two categories:data perspective and model perspective.The data-based method is mainly to tap more available information to alleviate the problem of cold startup.Recently,the method based on heterogeneous information network has attracted much attention because it can provide rich auxiliary information.However,in the absence of supervision data,the model cannot accurately use a large amount of information excavated by the heterogeneity.The vector of the generated users and content is not accurate,resulting in the decline in the accuracy of the recommendation results.For the above issues,this article is designed with a multi angle compared to the learning information mining model to effectively use the rich information on the heterogeneous diagram.This article first designs the characteristics of different information carried by multi-angle.The generation method can guide the model more accurately through the contrastive learning to tap the most related feature information of users and content,generate more accurate vector representation,and improve the accuracy of the prediction of the recommendation model.At the same time,some recent studies have also begun to alleviate the cold start problem from the perspective of the model.The meta-learning model can alleviate the cold start problem by sharing the knowledge of all users and building prior knowledge for users that can be used by all users.Most existing meta-learning algorithms assume that prior knowledge can be shared globally among all users,however,information sharing among users with hugely different interests is ineffective.And it will bring negative effects,which leads to a suboptimal solution of the model.This paper attempts fine-grained knowledge sharing only among users with similar interests.Specifically,this paper first proposes a user clustering method in the cold start scenario.Then,a transformation network is designed by using the user’s category information to transform the global initialization parameters learned by meta-learning methods into class-optimal initialization parameters.In this way,different degrees of information sharing can be carried out according to the similarity of users,reducing the negative impact of information sharing between users with large differences in preferences,and improving the performance of the meta-learning model.In order to verify the effectiveness of the meta-contrastive learning model algorithm,this article conducted experiments on two datasets,MovieLens and DBook.Compared to the base algorithm(MELU),under the user and item cold start scenario in MovieLens,mae decreases by 11.2%,RMSE decreased by 7.8%,Ndcg increased by 2.2%.Under the user and item cold start scenario in DBook,MAE decreased by 13.3%,RMSE decreased by 13%,Ndcg increased by 6.6%.The meta-contrastive learning model algorithm has shown significant improvement in multiple indicators across multiple datasets.