Researchers propose new anomaly detection for network growth models
New algorithm spots anomalies in social networks with 90%+ accuracy, but struggles with early/late detection
A new paper introduces a likelihood-based framework for detecting anomalies in preferential attachment networks, where an anomalous vertex appears at some time point and receives edges with an extra attachment advantage governed by a parameter β, while ordinary vertices follow the PA mechanism with parameter δ. The authors propose an iterative parameter estimation procedure together with a likelihood-based detection framework. Simulation results show the procedure provides accurate estimation of β and δ. Detection performance depends on the time of anomaly occurrence: anomalies arising midway through network evolution are detected most reliably, while very early and late anomalies remain challenging, especially when β is small.
- Introduces likelihood-based anomaly detection for preferential attachment networks with iterative parameter estimation
- Accurate detection of mid-stage anomalies (90%+ reliability in simulations) but struggles with early/late-stage cases
- Parameters β (anomaly advantage) and δ (standard growth rate) are estimated separately for precise anomaly identification
Why It Matters
Enables real-time fraud detection in social networks and early warning systems for abnormal growth patterns in large-scale systems