Within-Network Classification in Temporal Graphs

Abstract

Recent results indicate that static graph features might not be adequate to solve challenges in graphs involving a temporal dimension. We analyze several classification problems using already established temporal metrics, and we propose label-sensitive and recency-sensitive variants of these metrics that capture labeling information and additional temporal patterns in the data. We test all new and old metrics, and a baseline based on a standard disease-spreading model, using tuned off-the-shelf classifiers on 9 datasets of varying size and usage domain. Our experiments indicate that usage of label-and recency-sensitive metrics on real-world data provides more accurate results than static approaches and approaches based on temporal metrics alone.