Javier Raut
Back to Projects
schoolwork

DyMoN: A Post-Hoc Kinematic Approach for Detecting Anomalies in Dynamic Networks

A from-scratch kinematic model for anomaly detection in dynamic graphs. It tracks network evolution through physical motion metrics, calculating magnitude and direction scores—specifically isolating the tanh function for directionality. The model was benchmarked against DSEDN, JODIE, and STRIPE baselines to validate its motion-aware approach.

Technologies

PythonPytorchScikit-Learn NumpyPandas
Source

DyMoN: Dynamic Motion on Networks

The Origin: Back to the Drawing Board

The idea for DyMoN actually started in a cafe. We had just presented our original thesis concept, and we honestly didn't like where it was heading. We decided to scrap it and pivot. Sitting there, we started brainstorming a completely new angle for anomaly detection in dynamic graphs. Instead of looking at static changes, we realized we could track the evolution of network relationships like physical movement—and that's how the kinematic approach of DyMoN was born.

The Formulation: Scoring Dynamic Motion

To translate physical motion into network anomaly scores, we developed a specific mathematical formulation:

  • Magnitude & Direction: The DyMoN score integrates a magnitude-based component (calculated via the Euclidean norm of acceleration) and a direction-based component (derived from the cosine distance of velocity vectors).
  • Noise Reduction: We implemented Exponential Moving Average (EMA) smoothing to reduce noise and applied a hyperbolic tangent (tanh) gating mechanism to suppress stationary jitter.
  • Optimization & Efficiency: We established a tuning framework to optimize magnitude ($\alpha$) and direction ($\beta$) weights based on network topology. The algorithm runs with a strict linear time and space complexity of $O(N \times T \times D)$.

The Methodology: A From-Scratch Implementation

Rather than just modifying existing architectures, we implemented DyMoN entirely from scratch. To rigorously evaluate our custom kinematic approach, we utilized established architectures—specifically DSEDN, JODIE, and STRIPE—as our baseline models for benchmarking. Within our custom build, we integrated our specific scoring metrics, ensuring the tanh-based direction score calculation remained strictly separate from the overall DyMoN anomaly score.

The Result: Validation

Pivoting our topic in that cafe paid off. We proved that treating network changes as dynamic motion is a highly viable and computationally efficient way to detect anomalies.

  • Colloquium Success: We presented our research at the CS180 Student Research Colloquium on March 3, 2026, and took home the Best Presentation Award.
  • LAMBIGIT 2026: Our work was also accepted for the undergraduate poster presentation at the Northern Mindanao Research and Innovation Summit (April 28–30, 2026).