Community Embeggings with Bayesian Gaussian Mixture Model and Variational Inferrence for Recommendations

Graphs, such as social networks, emerge naturally from various real-world situations. Recently, graph embedding methods have gained traction in data science research.[17] Recommender systems are used in a wide range of business applications and are essential for online e-business models to survive and strive in the contemporary market. Using graph embeddings for recommendation tasks, have the possibility of improving upon recommender systems, because of data compression, their feature vector format, and sub-quadratic time complexity. The graph and community embedding algorithm ComE aims to preserve first-, second- and higher order proximity.[14] ComE requires prior knowledge of the number of communities K. In this paper, ComE is extended to utilize a Bayesian Gaussian mixture model with variational inference for learning community embeddings (ComE BGMM+VI). ComE BGMM+VI takes K as the maximum number of communities and drops components through a trade-off hyperparameter. Graph and community embeddings generated with ComE BGMM+VI are used to build a recommender system for friend suggestions. Recommendations are evaluated by the top-N hit-rate. A friend suggestions recommender system with a top-10 leave-one-out hit-rate of 43% is presented.