Social Recommendation Based On Graph Representation Learning

The age of big data speeds up the generation of information,and mankind has entered the information age with extremely sufficient information.While enjoying the convenience brought by the information age,people are also troubled by the problem of information overload.This has prompted the development of recommendation techniques.Although recommendation technology has developed rapidly in the past decades,there are still many problems to be solved,such as cold start and long tail problems caused by sparsity of preference information.The rapid growth of social networks offers sufficient social relations between users,and social network provide reliable data and information for recommendation.Social recommendation becomes research hotspot in recent years.Most of existing social recommendation methods are based on matrix factorization model and cannot mine social relations sufficiently.Consequently,those methods still suffer from cold-start problem,especially in users with few social relations.The graph representation learning algorithm has natural advantage on mining social relations.Therefore,this paper focus on graph representation learning based social recommendation.The work is as follows:Firstly,the social influence of users in the network is measured from the global and local views.The relation strength between two users is defined based on the learnt global and local social influence.Differ from existing social recommendation methods,we do not assume that all social relations are effective and several social noisy exists in social networks.Consequently,the user relation network with potential similar taste is constructed through observed user preference information.Consequently,user collaborative network with more reliable social relations is obtained by taking advantage of both the learnt user relation network and social network.Secondly,the low-dimensional user representations are obtained by performing random walk algorithm on.social graph.Based on different random walk sampling,the sets of similar users and dissimilar users are constructed.The final recommendation model is constructed on matrix factorization methods by combining two distance measures(maximizing similarity between similar users and differences between dissimilar users).Finally,a series of experiments have been conducted on four real-world data sets.Compared with the existing start-of-art social recommendation methods,the proposed collaborative social embedding method for matrix approximation(CoSEMA)obtains the best recommendation accuracy on all users and cold start users.Meanwhile,the experiments conducted on different user group with different social degrees demonstrates that CoSEMA has the ability to mining social relation more sufficiently and outperforms the existing social recommendation method on users with few social relations...