Transforming temporal-dynamic graphs into time-series data for solving event detection problems

Abstract

Event detection on temporal-dynamic graphs aims at detecting significant events based on deviations from the normal behavior of the graphs. With the widespread use of social media, many real-world events manifest as social media interactions, making them suitable for modeling as temporal-dynamic graphs. This paper presents a workflow for event detection on temporal-dynamic graphs using graph representation learning. Our workflow leverages generated embeddings of a temporal-dynamic graph to reframe the problem as an unsupervised time-series anomaly detection task. We evaluated our workflow on four distinct real-world social media datasets and compared our results with the related work. The results show that the performance depends on how anomalies deviate from normal. These include changes in both size and topology. Our results are similar to the related work for the graphs where the deviation from a normal state of the temporal-dynamic graph is apparent, e.g., Reddit. On the other hand, we achieved a 3-fold improvement in precision for the graphs where deviations exist on size and topology, e.g., Twitter. Also, our results are 20% to 5-fold better even if we introduced some delay factor. That is, we beat our competition while detecting events that occurred some time ago. As a result, our study proves that graph embeddings as time-series data can be used for event detection tasks.