Detecting Concept Drift in Classification Over Streaming Graphs

Detecting concept drift in data streams has been widely studied in the data mining community. Conventional drift detection methods use classifiers’ outputs (e.g., classification accuracy, error rate) as indicators to signal concept changes. As a result, their performance greatly depends on the chosen classifiers. However, there is little work on addressing concept drift in graph-structured data. In this paper, we present a Graph Entropy-based Method (GEM) to effectively detect concept drift in graph streams. Contrary to many related works, we investigate the intrinsic properties of data (i.e., subgraph distribution w.r.t. class membership), instead of monitoring classification outputs. This method can be combined with any graph stream classifier to facilitate classification on non-stationary graph streams. Our approach is combined with several graph stream classification algorithms and tested on synthetic and real-world graph data streams. The experimental results demonstrate the advantage of our method in detecting concept drift as well as improving classification performance.