GAN-SOM: A clustering framework with SOM-similar network based on deep learning

摘要

Since an increasing amount of data is generated and collected in real life, clustering is more frequently applied to process these unlabeled data in practical problems. Due to the simple similarity measure of conventional clustering methods, they are unable to achieve good performance on current big data. With the popularity of deep learning, deep clustering has been developed in recent years and obtained remarkable results. However, they have complex architecture and consume numerous computational resources, which goes against the migration to edge devices. Therefore, methods with low cost are required to satisfy edge computing, which is the trend of development. In this paper, we propose GAN-SOM as a new architecture for clustering based on deep learning. A SOM-similar network is designed to simultaneously implement encoding and clustering purposes on data samples, which is jointly trained with a GAN to optimizes a new defined clustering loss. We also utilize self-attention mechanism and spectral normalization in the GAN architecture to enhance effects of generated data, which aims to achieve better clustering results. The experimental results compared with other clustering baselines with deep learning verify that our method maintains high clustering metrics while saving computational cost significantly.