Conditional Local Filters with Explainers for Spatio-Temporal Forecasting
share
Spatio-temporal prediction is challenging attributing to the high nonlinearity in temporal dynamics as well as complex dependency and location-characterized pattern in spatial domains, especially in fields like geophysics, traffic flow, etc. In this work, a novel graph-based directed convolution is proposed to capture the spatial dependency. To model the variable local pattern, we propose conditional local filters for convolution on the directed graph, parameterized by the functions on local representation of coordinate based on tangent space. The filter is embedded in a Recurrent Neural Network (RNN) architecture for modeling the temporal dynamics with an explainer established for interpretability of different time intervals' pattern. The methods are evaluated on real-world datasets including road network traffic flow, earth surface temperature & wind flows and disease spread datasets, achieving the state-of-the-art performance with improvements.
READ FULL TEXT
