Local event forecasting and synthesis using unpaired deep graph translations

Abstract

Local rare event forecasting and synthesis on networks are highly useful for emergence management. For example, synthesizing traffic congestion and disease diffusion over the road network and disease-contact network respectively of specific geo-locations is highly important for transportation planning and disease outbreaks intervention. This task requires to learn how the events of congestion or disease “translate” the graph patterns from source mode (e.g., without event) to target mode (e.g., with event) based on historical data for some locations. Then it needs to apply such “translation” upon a source-mode graph pattern in a new location’s network, in order to estimate and foresee what it will look like in target-mode in this location. Such task is called graph translation, which is an analogy and generalization to image and text translation. Similar to the situations in image and text translation, paired training data, which consists of pairs of source-mode graph and its corresponding target-mode, will usually not be available. In this work, we propose an approach for learn the translation of graphs from source-mode to target-mode such that the generated target-mode is indistinguishable from the distribution of the real target-mode using an adversarial loss. Because there is no paired training data, we also learn an inverse translation from target-mode to source-mode and couple these two translation mappings through cycle consistency loss. Extensive experiments on both synthetic and real-world application data demonstrate that the proposed approaches is capable of generating graphs close to real target graphs. Case studies on the synthesized networks have also been illustrated and analyzed to show the reasonableness of the generated target-mode graphs.