Business Context

AeroWave Networks is tuning an OFDM-based wireless access product used in dense urban deployments. The radio team wants a lightweight ML model that predicts whether a transmission interval will meet quality targets so the scheduler can choose robust modulation and coding settings.

Dataset

You are given a labeled dataset of OFDM transmission intervals collected from field trials and lab replay. Each row represents one 10 ms interval for a single user equipment (UE), with engineered features derived from the PHY/MAC stack.

• Size: 240K intervals, 40 features
• Target: Binary label — link_quality_good = 1 if block error rate (BLER) < 10% and throughput exceeds the service threshold; else 0
• Class balance: 68% good, 32% poor
• Missing data: 8% missing in mobility fields, 3% missing in some subcarrier summaries due to telemetry dropouts

Success Criteria

A good solution should achieve ROC-AUC >= 0.88 and F1 >= 0.78 on a held-out test set, while keeping inference latency low enough for near-real-time scheduling support.

Constraints

• Inference budget: under 5 ms per batch of 100 intervals
• Model should be explainable to RF engineers
• Retraining should be feasible weekly as network conditions shift
• Avoid leakage from future intervals of the same session

Deliverables

1. Train a binary classifier for OFDM link-quality prediction.
2. Explain why OFDM-related features matter and how they connect to wireless performance.
3. Show preprocessing for mixed numeric/categorical data and missing values.
4. Use a leakage-safe validation strategy.
5. Report metrics, threshold choice, and top feature importances.