HONEM:Learning Embedding for Higher Order Networks

Representation learning offers a powerful alterna- tive to the oft painstaking process of manual feature engineering, and as a result, has enjoyed considerable success in recent years. This success is especially striking in the context of graph mining, since networks can take advantage of vast troves of sequential data to encode information about interactions between component of a complex system. While there has been a recent spate of research on representation learning on networks, all such methods operate on the first-order networks, ignoring the potential non-Markovian interactions between the nodes. To address this challenge, this paper presents a network embedding method that leverages the complex interactions encodes in higher-order networks (HON) for inferring the higher-order neighborhood information. We demonstrate that the higher- order network embedding (HONEM) method learns the higher- order patterns in HON, and outperforms other state-of-the-art methods in node classification, network re-construction, link prediction, and visualization.