Problem
Traffic streams contain strong temporal patterns, noise, and event-driven spikes that make anomaly detection easy to overfit.
Project
PMML Project: Traffic Anomaly Detection in Toronto
Anomaly detection on urban traffic streams comparing probabilistic GLM and sequence-model (GRU) approaches.
Approach
I compared a probabilistic GLM baseline against a GRU sequence model, then analyzed how calibration and divergence-based scoring behaved under real traffic variation.
Result
The work clarified when simpler probabilistic models remain competitive and where sequence models help once temporal context matters.
Focus
This project was less about chasing one model score and more about understanding what different anomaly detectors assume.
What I compared
- Poisson/GLM-style probabilistic modeling for interpretable traffic count behavior.
- GRU-based sequence modeling for temporal dependency capture.
- Divergence-oriented anomaly scoring to inspect shifts in learned behavior.
Main lesson
If the anomaly score is poorly calibrated, a more complex model can still produce noisy operational signals. Evaluation had to stay tied to practical alert usefulness, not just loss curves.