Adaptive Drift Intelligence | Team Hack Tuah

🏆 Winning Solution: Singtel x National AI Student Challenge 2026
🏆 Best-performing submission across all teams nationally.

A schema-agnostic pipeline for robust churn prediction under distribution shift. Diagnoses drift per feature, applies targeted mitigation, and adapts the model to the test distribution through iterative self-training, all within a fixed model and 10-minute runtime constraint.

0.893 AU-PRC on the public evaluation dataset — up from a baseline of 0.723 (+23.5% relative improvement).

---

Quick Start

# Install dependencies
pip install -r requirements.txt

# Run the pipeline
python ./src/main.py --train_data_filepath <train_csv_path> --test_data_filepath <test_csv_path>

Example with public data:

python ./src/main.py --train_data_filepath data/train.csv --test_data_filepath data/test.csv

Estimated runtime: ~30 seconds on the public dataset (70K rows, 42 features). Scales to under 10 minutes on datasets up to 10M rows and 500 features.

---

Outputs

The pipeline produces all required outputs automatically:

Output	Location	Description
Drift summary table	Console	Per-feature drift type, description, and mitigation applied
Runtime	Console	Total execution time in seconds
AU-PRC metrics	Console	Train and test set AU-PRC
Predictions preview	Console	Head of predicted probabilities
prediction.csv	Root directory	CustomerID + probability_score for all test rows
model.joblib	Root directory	Trained LightGBM model

---

Dashboard

Start the dashboard server, then use the browser GUI to upload CSVs and run the pipeline:

python dashboard/server.py

Open http://localhost:5050 — upload train/test CSVs, click ▶ Run Pipeline, and results appear automatically. No terminal interaction needed.

The dashboard includes two tabs:

• Overview: metrics, drift severity chart, pipeline ablation, score distribution, drift summary table
• All Predictions: searchable, sortable table of all predictions with risk levels and CSV download

---

Project Structure

.
├── src/                        # Solution source code
│   ├── main.py                 # Entry point — pipeline orchestration
│   ├── drift_detection.py      # Phase 1: per-feature drift diagnostics
│   ├── encoding.py             # Phase 2: feature encoding & mitigation
│   ├── adaptation.py           # Phase 3+4: temporal weighting, self-training
│   └── utils.py                # Shared constants, sampling, stats
├── dashboard/                  # Interactive web dashboard
│   ├── index.html              # Dashboard UI
│   └── server.py               # Lightweight HTTP server
├── prediction.csv              # Model predictions on public test set
├── model.joblib                # Trained model
├── requirements.txt            # Python dependencies
└── README.md                   # This file

---

How It Works

1. Diagnose: Each feature is independently tested for distribution shift (KS/χ²), concept drift (temporal correlation stability), and format changes (new categories). Features are classified into drift types: none, covariate, concept, format, mixed, or severe.

2. Mitigate: Shifted numerics are quantile-mapped. All categoricals are target-encoded with case normalisation (resolves format changes implicitly). Severe mixed-drift features are dropped. Concept drift triggers temporal weighting.

3. Adapt: Iterative self-training progressively adapts the model to the test distribution using high-confidence pseudo-labels. Budget-constrained, safety-gated, with prediction averaging for stability.

---

Requirements

• Python 3.10+
• LightGBM 4.6.0
• scikit-learn ≥ 1.3.0
• pandas ≥ 2.0.0
• numpy ≥ 1.24.0
• joblib ≥ 1.3.0

---

Team Members

• Garv Sachdev
• Yoong Hong Jun, Nicholas
• Glynis Looi Xin Lin
• Jan Chen Jie
• Ronav Pattanaik

---