Title:Graph-Survival: A Survival Analysis Framework for Machine Learning on Temporal Networks

Abstract:Continuous time temporal networks are attracting increasing attention due their omnipresence in real-world datasets and they manifold applications. While static network models have been successful in capturing static topological regularities, they often fail to model effects coming from the causal nature that explain the generation of networks. Exploiting the temporal aspect of networks has thus been the focus of various studies in the last decades.
We propose a framework for designing generative models for continuous time temporal networks. Assuming a first order Markov assumption on the edge-specific temporal point processes enables us to flexibly apply survival analysis models directly on the waiting time between events, while using time-varying history-based features as covariates for these predictions. This approach links the well-documented field of temporal networks analysis through multivariate point processes, with methodological tools adapted from survival analysis. We propose a fitting method for models within this framework, and an algorithm for simulating new temporal networks having desired properties. We evaluate our method on a downstream future link prediction task, and provide a qualitative assessment of the network simulations.