From Structure to Strength: A Deep Learning Approach to Edge Weighting in Graph-Based Community Detection

Abstract

Complex networks analysis requires a full comprehension of the interconnection between the nodes and how these interconnections give rise to community structures. This is done by finding clusters of nodes that are well connected within the cluster but relatively poorly connected with the rest of the network. Out of the many methods conceived to make this task easier, edge weighting has become an indispensable tool, allowing one to quantify the strength of connection in terms of the importance or level of interaction between the nodes. In this work, we propose a new mechanism of edge weighting based on representation learning using autoencoders. The model is trained using the network's adjacency matrix to learn latent structural aspects of the nodes in the network, which are then used as a basis of calculating pairwise similarities using an amplified Gaussian similarity function. As a methodology, this allows us to capture some of the more subtle structural patterns that cannot be identified through traditional local heuristics. Due to the experimental results, we show that our method can give the same performance in a community detection task to popular similarity-based weighting strategies such as Jaccard, cosine similarity, and the Adamic-Adar index, thus proving that learned representations may find success in these kinds of collaborations.