Parallel stacked autoencoder with particle swarm optimization for community detection in complex networks

Community detection is one of the long standing and challenging tasks in the field of Complex Networks (CNs). Recently, deep learning is one of the promising community detection methods, which can learn effectively low-dimensional representation of CNs. However, the existing methods have major drawbacks in terms of local minima and slow convergence, since they use Gradient Descent Backpropagation algorithm (GDBP). This reduces the performance of community detection in terms of effectiveness and efficiency. To overcome these drawbacks, this paper introduces a new parallel deep learning model based on Metaheuristic (MH) algorithm instead of the GDBP algorithm. To be specific, a new parallel stacked autoencoder (SAE) based on particle swarm optimization (PSO) is developed for feature learning and community detection in CNs. The PSO algorithm uses a multi-objective fitness function that includes the standard loss function (i.e., MSE) of the autoencoder and the modularity function to guide SAE optimization and improve community detection performance. In addition, an efficient distributed parallel implementation is proposed to improve the efficiency and scalability of the SAE-based PSO method. The parameter settings of PSO such as features-dimension and number of particles, are tuned and studied to observe their implications on community detection performance. We conducted an experiment comprising datasets of 10 real-world networks to evaluate the proposed method in different parameter settings. The results demonstrated that the SAE-based PSO method is promising and provides a competitive performance against state-of-art methods in community detection. Furthermore, the results showed that the parallel implementation of the proposed method could improve efficiency with three or greater orders of speed.