A Filtered Gradient Descent Clustering Method to Recover Communities in Attributed Networks

An attributed network is a network in which in addition to the network structure each node is associated with a set of attributes. Community detection in such networks involves recovering the clusters by simultaneously using the network data and the node attributes. This research proposes a generic objective function and adopts the gradient descent (GD) approach to recover clusters in attributed networks. Our straightforward adoption of GD, even in its improved version, such as the adaptive moment estimation method, may encounter "bad sequences" of objects and converge to points that are far from optimal. To tackle this issue, we introduce a special "filter" mechanism, which culls potentially misleading objects. We established the theoretical foundation of the proposed filtering mechanism and the corresponding clustering method, and empirically evaluated and compared the performance of our proposed methods using synthetic and real-world datasets, including one new real-world dataset. Our results show that our proposed filter mechanism does improve results significantly and makes it competitive versus state-of-the-art algorithms.