Research On Collaborative Filtering Recommendation Algorithm Based On Neural Network

With the rapid development of Internet technology and the rapid emergence of various Internet businesses,people’s lifestyles have changed dramatically.As a result,the data on the internet has shown explosive growth.The massive amount of data has caused difficulties for users to quickly obtain useful information,which is the phenomenon of "information overload".Although the search engine alleviates the problem of "information overload" to a certain extent,it cannot meet the diversified needs of users,so the recommendation system technology comes into being.Collaborative filtering is a mainstream recommendation algorithm in recommendation systems,which is based on the user’s historical behavior to predict the rating of items.In recent years,thanks to the development of neural network of the fields of image processing,natural language processing and artificial intelligence,combining neural networks with traditional recommendation algorithms to perform recommendation tasks has become a mainstream approach.In neural collaborative filtering,neural networks can better capture non-linear and non-trivial user-item relationships,the design of neural collaborative filtering recommendation algorithm mainly faces two problems: how to learn the representation of users and items and how to model the interaction between users and items.The work in this article is dedicated to the analysis of collaborative filtering algorithms based on neural networks,and the main work consists of the following two parts.1.In order to better model user-item interaction,this paper proposes a outer and hadamard product-based joint convolutional collaborative filtering.Firstly,a user-item rating matrix is constructed by explicit ratings and implicit feedback.Explicit ratings are not only used to train the recommendation model parameters,but also used to represent the real preferences of users and the original attributes of the items,different ratings represent different degrees of preferences of users to the items.Then,with this matrix as the input,the deep neural network is used to extract the nonlinear low-dimensional features of users and items,and the outer product and hadamard product are used to model the pair-wise correlation between users and items and the correlation between users’ features,so as to generate a two-dimensional interaction map with more expressive power and semantic rationality.Finally,convolutional neural networks are used to learn higher-order correlations between users and items in the interaction map.2.A squeeze and excitation networks-based factorization machines collaborative filtering is proposed.In order to better learn the representation of users and items,sparse high-dimensional rating vectors are mapped to low-dimensional dense vectors through a full-connection layer.Then,a more expressive 2D matrix is generated by performing separate self-outer product operations on the user vector and the item vector.Finally,the user and item representation is learned from 2D matrix by squeeze and excitation networks,which can increase the weight of important features and reduce the weight of non-important information features.In terms of user-item interaction modeling,in order to consider a more comprehensive user-item interaction relationship,not only the fully connected neural network is used to model the high-order interaction between users and items,but also a factorization machines is designed,which can model the low-order interaction between users and items and further improve the overall recommendation performance.