Multi-task Neural Collaborative Filtering Model For Personalized Recommender System

With the rapid development and popularity of the Internet,the amount of information in the network is exploding,and it is difficult for users to filter the content they are interested in,which leads to the problem of information overload.A recommender system is an important tool to solve information overload,and collaborative filtering is the most widely researched and applied algorithm in the recommender system.In recent years,recommender systems have been integrating deep learning to improve performance.Neural network-based recommendation algorithms have stronger characterization ability,which can dig out users’ preferences from user’s behavior records and alleviate the scalability problem of traditional recommendation algorithms.There are multiple ways to record the user’s interaction with items,and the records of different ways all contain rich and valuable information of users’ preferences.Therefore,how to fuse these valuable information to improve the performance of recommender systems is a valuable research topic.At the same time,the number of users and items in a recommender system is often very large,and it has become a hot topic in the field of recommender systems to use historical interaction data to accuratelyextract embedding vectors of users and items,so as to improve the convergence speed and the performance of recommendation models.In this paper,we study the recommendation methods based on multi-task neural collaborative filtering,and explore the methods of extracting embedding of users and items to improve the recommendation performance of the model.The specific work is as follows.(1)A multi-task neural collaborative filtering model based on deep fusion of explicit and implicit ratings is proposed.Explicit and implicit ratings are the two most common types of ratings used in recommender systems.An explicit rating refers to the specific rating of an item given by a user,which can intuitively indicate how much the user likes the item.Implicit ratings are the records of whether there is an interaction behavior between the user and the item,and can provide a comprehensive representation of the range of possible interests of the user.This model deeply fuses the two types of rating information in terms of both the network structure and the loss function.In the aspect of the network structure of the model,two auxiliary tasks are designed to learn the explicit and implicit user-item rating features respectively,and the two feature vectors are concatenated into the main task network,after which the prediction rating is obtained through fully connected layers.In the aspect of the loss function of the main task,we construct weight terms based on the explicit ratings and design a weighted binary cross-entropy loss function,thus incorporating the explicit ratings into the binary cross-entropy.This is equivalent to adding an attention mechanism to the loss function,so that the recommendation model devotes more attention to the highly rated records during the training process and obtains better recommendation performance.(2)A multi-task collaborative filtering model based on User&Item2vec and deep fusion of explicit and implicit ratings is proposed.To solve the problems of too many parameters,high computational cost and bad optimization of the main structure in the training process of the embedding layer of the end-to-end model,User&Item2vec is designed to first extract the embedding vector of users and items,and then input the user embedding and item embedding to the multi-task collaborative filtering model with deep fusion of explicit and implicit rating.The User&Item2vec method is modeled based on historical user-item interactions that uses a single hidden layer network structure to quickly train and extract embeddings of users and items.Experimental results on real-world datasets with three different sizes show that extracting embedding vector of users and items based on User&Item2vec and using a multitask collaborative filtering model leads to better performance.(3)A multi-task collaborative filtering model based on User&Item2vec and generalized matrix factorization is proposed.To further improve the accuracy of the recommendation results,a generalized matrix decomposition module is added to the multi-task collaborative filtering model,so that the model has both "generalization ability" and "memory ability" for the input features.At the initialization stage of the model parameters,the commonly used parameter initialization method which does not contain a priori information,such as random initialization,is not adopted.Instead,we initialize the embedding layer using the embedding vectors extracted by User&Item2vec,which can speed up the convergence of the model and improve the accuracy of the recommendation results.The experimental results show that adding the generalized matrix factorization module can enhance the memory capability,and at the same time,using the a priori information for parameter initialization can further improve the performance.