Benchmarking Edge Regression on Temporal Networks

Abstract

Benchmark datasets and task definitions in temporal graph learning are limited to dynamic node classification and future link prediction. In this paper, we consider the task of edge regression on temporal graphs, where the data is constructed from sequence of interactions between entities. Upon investigating graph benchmarking platforms, we observed that the existing open source datasets do not provide the necessary information to construct temporal edge regression tasks. To address this gap, we propose four datasets that naturally lend themselves to meaningful temporal edge regression tasks. We evaluate the performance of a set of method based on popular graph learning algorithms in addition to simple baselines such as vertex-based moving average. Processed versions of proposed datasets are accessible through this repository: huggingface.co/cash-app-inc.