Study On Image Recommendation Methods Based On Triple Collaborative Filtering

With the rapid development of Internet technology and the increasing popularity of various electronic devices,as well as the rapid spread of social media,it is more convenient to generate and spread images.According to statistics,nearly 80%of big data on the Internet is image data,which leads to serious visual information overload.Image recommendation technology,as a supplementary of traditional image retrieval technology,can effectively alleviate this problem and play an increasingly important role in various multimedia applications.The interactive data between users and images(usually represented as UI matrix)is an important basic resource for building image recommendation system.However,compared with the number of users and images,the interactive information between users and images is extremely sparse,which is one of the main challenges faced by image recommendation technology.At present,a mainstream solution is to use the side information of users and images to assist the recommendation task,so as to alleviate the problem of the sparseness of user-item interactions and improve the recommendation performance.However,in most image search engines,only the unilateral side-information(i.e.image visual features)is available,and it is difficult to obtain the user's personal information.Therefore,for the image recommendation task,a tough issue is how to enhance both user and image representations using unilateral image visual information.This thesis is based on solving the above problems,and the work is as follows:Firstly,a Tri-CF framework is proposed,which extends the classical bayesian personalized ranking model with two smooth terms related to user and image respectively,so that the user-user and image-image similarity can be taken into account when modeling user-image preference.This framework has high flexibility and expansibility,and can easily integrate all kinds of user and item side information into the similarity calculation,so as to improve the representation of user and image.In addition,based on the interaction history between the user and the image,a side-information transfer strategy is proposed to improve the implicit representation learning of both the user and the image by using the unilateral visual information of the image.Concretely,first of all,the hybrid similarity matrix of the image is calculated by using the vision of the image and the click log of the user.Then spectral clustering algorithm is used to group the images adaptively.Based on the low-dimensional subspace of image clustering,the BOW-like representation of the user is obtained,and the similarity matrix of the user's subspace is calculated.In particular,hybrid image similarity matrix and subspace user similarity matrix are respectively used for image smooth and user smooth in the Tri-CF framework.In this way,the representations learning of user and image is improved by using only single side-information.Finally,a large number of experimental results show that the proposed scheme can use visual information to improve the representation of users and images at the same time,and its performance is much better than the existing similar methods.Finally,by comparing the proposed algorithm with its degradation algorithms and the existing classical algorithms,it is proved that the proposed scheme can improve the representation of users and images simultaneously by using visual information,and its performance is much better than that of the existing similar methods.