Research On Cold-start Recommendation System Based On Meta-learning

With the rapid development of mobile network and the update and iteration of storage hardware,the scale of Internet data is increasing day by day.How to enable Internet users to quickly seek the information they are interested in in the diverse data ocean has become one of the challenging problems in the field of data mining.As a technical tool to quickly locate the user’s interested information,the research of personalized recommendation system is of great significance to solve this problem.However,when a new user or item just enters the system,due to the lack of relevant data,the recommendation system is difficult to play a role compared with ordinary users or items,and the "cold start problem" also appears.In view of the strong sparse data processing capability,meta-learning paradigm is gradually becoming a key technology to alleviate the cold-start problem.By learning the meta-global parameters to initialize the recommendation model,the cold-start problem can be alleviated to a large extent,which can help the recommendation system to carry out effective personalized recommendation.Based on meta-learning paradigm,this paper studies the cold-start problem of recommendation system,and the work is mainly divided into the following three points:(1)In view of the fact that most studies only consider data or models from a single level,it is difficult to effectively alleviate the problem of cold-start.Therefore,in this paper,a memory-augmented meta-learning on meta-path for fast adaptation cold-start recommendation is proposed by combining meta-path and memory-augmented meta-optimization at both data and model levels.The model introduces meta-path at the data level to capture semantics and dig out deeper relationships between users and items.Model-level semantic adaptors learn unique semantic priors brought by all meta-paths and update each user’s preference by learning various semantic priors for the recommendation model.At the same time,to solve the local optimization problem,a special semantic storage unit is designed by the model to provide personalized bias items for initialization model parameters.Finally,semantic and task adaptation in large parameter space of neural network is realized by single-step highdimensional gradient descent.(2)Previous studies strongly rely on collaborative filtering technology when integrating multiple semantics,and fine-tuned structures are difficult to adapt to specific new users,which may lose the relevance of some features or items,and also lead to performance loss of multisemantic fusion.To solve the above problems,this paper uses a meta-learning-based approach to generate an ideal initial embedding for the user’s semantic context and item.Specifically,user-related semantics and items are introduced into user-specific tasks,and new semantic embedding vectors or new item embedding vectors are generated by constructing multiple fully connected neural networks and using previous semantics and item priors,and parameters are updated by meta-optimization method.(3)In order to comprehensively evaluate the proposed method,two commonly used real datasets DBook and Movielens-1M are selected as the benchmark datasets,and experiments are carried out between the proposed method and other benchmark models in three cold start scenarios and conventional recommendation scenarios.The performance of the method on evaluation indexes RMSE,MAE and NDCG@5 was tested.Experimental results verify the effectiveness of the proposed memory-augment meta-learning on meta-path for fast adaptation cold-start recommendation.