Random Walk based Conditional Generative Model for Temporal Networks with Attributes

Stratis Limnios, Andrew Elliott, Mihai Cucuringu, Gesine Reinert

Keywords: Graph, GAN, synthetic data, Temporal Graphs, Graph attributes

Abstract: We propose a novel method for graph time series generation with node and edge attributes. As graph representations for complex data become increasingly popular, we encounter many time series of graphs with temporal and attribute dependencies, such as communication networks, daily bike rentals or bank transactions. However, the analysis of such graphs can be impeded by privacy or data protection issues, calling for synthetic network time series which serve as surrogate of the observed time series. There are many methods to construct networks, for example transformer-based graph models such as TagGen, but it is harder to create data which are faithful to the observed dependencies, including those of attributes across time. Moreover, tabular data generation methods, such as GANs, fail to emulate complex graph structures. To circumvent these limitations, we introduce CTWalk which combines and leverages the strengths of both methods, by coupling TagGen that learns the distribution of temporal random walks over the input data, and a conditional tabular GAN (CTGAN) that captures the time dependence of the features. CTWalk is able to mimic edge weight distributions, node labels, and temporal dependencies of the data.