OCC-NCS: One-Class Classification through Network Community Structure Analysis

One-class classification is essential for effectively detecting anomalies and rare events across various applications, particularly in situations where obtaining labeled data for these rare events is both challenging and costly. In this study, we develop a novel one-class classifier based on the community structure of complex networks. Instances of the considered class are used to form a complex network, where nodes represent data points and edges denote similarity or distance. The Label Propagation Algorithm (LPA) is employed to identify communities within this network, grouping closely related instances. After communities are formed, new data points are classified by assessing their membership in these established communities. If a new point has a sufficient number of neighbors within a community, it is classified as a member of that community; otherwise, it is labeled as an outlier or abnormal event if it fails the membership assessment across all communities. This community-based classification approach enhances the robustness of rare event detection by leveraging the structural relationships among similar instances within the dataset, as evidenced by benchmarking with established one-class classification methods.