Community Preserving Network Embedding Based on Memetic Algorithm

Abstract

Network embedding aims to embed network nodes into a low-dimensional and continuous vector space, which can benefit various downstream network analysis tasks. As it is an emerging topic in recent years, a variety of methods have been proposed to learn representations by preserving a network topology structure. However, it still remains challenging to incorporate a community structure into network embedding, which is ignored by most of the methods. In this paper, we present a novel memetic algorithm for network embedding, which is termed as MemeRep. As a matter of fact, the community structure is preserved by optimizing the modularity density. In our methods, genetic algorithm is adopted to optimize a population of solutions, and a problem-specific local search procedure with the two-level learning strategies is designed to accelerate the optimization process. The first-level learning strategy enables each node to learn from its neighbors, while the second-level learning strategy expands the learning area, which enables each node to learn from communities. Experiments on real-world and computer-generated networks show that the proposed algorithm outperforms several state-of-the-art methods in visualization, node classification, and community detection.