Evaluating the Community Structures from Network Images Using Neural Networks

摘要

Communities or clusters in a network unveil important structures of a physical or social system. Suppose a Google image search for "complex network" returns thousands of network images. Can we infer the community structures in the original networks just from these images? Traditional community detection algorithms will fail in this case because they do not work with images. We developed an approach where convolutional neural networks (CNNs) are trained to reveal community information from images of complex networks. We trained three CNNs with images of simulated networks having ground truth communities. The training process uses state-of-the-art community detection, graph drawing, and deep learning training algorithms. The trained networks are then used to predict the number of communities and the modularity score (a measure of community structure of networks) for real-world networks. We formulated these two tasks as a classification and a regression problem and used appropriate loss functions for them. The CNN models can attain test accuracy of 81% and 33.3% for simulated and real networks, respectively. This result is statistically significant as can be seen by Spearman's rank correlation of 0.77 with a p-value of 6.2 × 10~(-13) for real-world networks.