Business Context

OptiWave Networks operates a national DWDM transport backbone with 9,000+ optical spans and hundreds of ROADMs. The network operations team wants a model that predicts whether a wavelength service will experience a photonic-layer incident in the next 24 hours so they can prioritize preventive maintenance and rerouting.

Dataset

You are given telemetry aggregated at the channel-hour level from DWDM shelves, ROADMs, amplifiers, and photonic controllers.

• Size: 1.2M channel-hour records across 14 months, 54 features
• Target: Binary label indicating whether the channel had a photonic-layer incident within the next 24 hours
• Class balance: 2.9% positive, 97.1% negative
• Missing data: 8% missing in OSNR/Q metrics, up to 20% missing for some vendor-specific counters

Success Criteria

A good solution should achieve recall >= 0.75 on incidents while maintaining precision >= 0.35 for the alert queue. The model should also provide feature-level explanations that help operations engineers understand likely root causes.

Constraints

• Batch scoring every 15 minutes; per-record inference should stay under 20 ms
• Predictions must be explainable enough for NOC and photonic engineering teams
• Avoid leakage from future alarms, tickets, or post-incident remediation data

Deliverables

1. Build a binary classification pipeline for next-24-hour incident prediction
2. Explain feature engineering choices tied to DWDM/ROADM behavior
3. Choose evaluation metrics appropriate for severe class imbalance
4. Propose a validation strategy that respects time ordering and repeated channels
5. Describe how you would deploy, monitor, and retrain the model in production