No Community Detection Method to Rule Them All!
Abstract
Community detection is a core tool for analyzing large realworld graphs. It is often used to derive additional local features of vertices and edges that will be used to perform a downstream task, yet the impact of community detection on downstream tasks is poorly understood. Prior work largely evaluates community detection algorithms by their intrinsic objectives (e.g., modularity). Or they evaluate the impact of using community detection onto on the downstream task. But the impact of particular community detection algortihm support the downstream task. We study the relationship between community structure and downstream performance across multiple algorithms and two tasks. Our analysis links community-level properties to task metrics (F1, precision, recall, AUC) and reveals that the choice of detection method materially affects outcomes. We explore thousands of community structures and show that while the properties of communities are the reason behind the impact on task performance, no single property explains performance in a direct way. Rather, results emerge from complex interactions among properties. As such, no standard community detection algorithm will derive the best downstream performance. We show that a method combining random community generation and simple machine learning techniques can derive better performance