A multi-granular network representation learning method

Granular computing (GrC) as a problem-solving concept and new information processing paradigm is deeply rooted in human thinking, which has attracted many researchers to study it theoretically, and has gradually applied to data-driven problems. Network embedding, as known as network representation learning, aiming to map nodes in network into a low-dimensional representation, is a data-driven problem. Most existing methods are based on a single granular, which learn representations from local structure of nodes. But global structure is important information on the network and has been proven to facilitate several network analysis tasks. Therefore, how to introduce GrC into network embedding to obtain a multi-granular network representation that preserves the global and local structure of nodes is a meaningful and tough task. In this paper, we introduce Quotient Space Theory, one of the GrC theories into network embedding and propose a Multi-Granular Network Representation Learning method based on Quotient Space Theory (MG_NRL, for short), which can preserve global and local structure at different granularities. Firstly, we granulate the network repeatedly to obtain a multi-granular network. Secondly, the embedding of the coarsest network is computed using any existing embedding method. Finally, the network representation of each granular layer is learned by recursively refining method from the coarsest network to original network. Experimental results on multi-label classification task demonstrate that MG_NRL significantly outperforms other state-of-the-art methods.