Graph signal processing on dynamic graphs based on temporal-attention product

Ru Geng, Yixian Gao, Hong-Kun Zhang, Jian Zu

Abstract: Signal processing is an important research topic. This paper aims to provide a general framework for signal processing on arbitrary dynamic graphs. We propose a new graph transformation by defining a temporal-attention product. This product transforms the sequence of graph time slices with arbitrary topology and number of nodes into a static graph, effectively capturing graph signals' spatio-temporal dynamic evolution process. The temporal-attention product graph provides a solid mathematical foundation to model the time-dependent graph signal processes as martingales. The weighted adjacency matrix obtained by temporal-attention products is a block tridiagonal matrix, which has been extensively studied. Therefore, it is general and convenient to perform graph signal processing on this new static graph. We apply two real datasets to illustrate the effectiveness of spectral graph wavelet transform based on temporal-attention product. For one of the datasets with no graph structure, we learn the graph weights through a neural network.