Building a Graph-Based Recommender Using Community Embeddings
In this work, we explore the application of graph embedding to the design and development of a friend recommender system for the users of the social network. Graph embedding could be useful for recommendation tasks because of data compression, the feature vector format, and sub-quadratic time complexity of graph embedding. We suggest and study a ComE BGMM+VI algorithm that is essentially a proprietary modification of the ComE community embedding algorithm where Bayesian Gaussian mixture model and variational inference are used for community embedding and detection. Graph and community embedding generated with this algorithm are intended for the recommender system for social network friend suggestions. Experiments with prototype recommender were conducted on popular graph datasets of Zachary's Karate Club and Social Circles from Facebook. Generated recommendations were evaluated by the top-N hit-rate for users with at least 50 friends. A prototype recommender demonstrates a top-10 leave-one-out hit-rate of 43.6% and run-time optimized hit-rate of 32.9%.