HackNU 2026 — The Retention Architect

Higgsfield DS Challenge — predictive churn system that distinguishes Voluntary Churn (user chooses to cancel) from Involuntary Churn (payment failures).

3-class classification: not_churned (50%) / vol_churn (25%) / invol_churn (25%)

Data → Feature Engineering → Models → Stacking → Calibration → Threshold Tuning → Submission

Properties — subscription start date, country, temporal features (year 1067 synthetic dates parsed as strings)

Purchases — total spend, purchase types, purchase velocity

Transactions — dual approach for failed tx (no user_id):

• Approach 1: Fingerprint matching (card_brand + bank + country + funding) between failed and successful transactions, weighted 1/N for multi-user fingerprints
• Approach 2: Card risk metadata from successful transactions (is_prepaid, cvc_check, country_mismatch)

Generations (5.5GB, read in chunks) — temporal patterns per user:

• day01_ratio, burst_flag — captures "binge and leave" pattern for vol_churn
• 14 × active_day_X binary flags (daily retention)
• first_quit_day, max_active_streak
• Week1 vs Week2 trends (gen_trend, credits_trend)
• Recency features (days_since_last_gen)

Quizzes — onboarding signals (source, frustration, experience)

Class-specific interactions:

• prepaid_x_failure, payment_risk_score → invol_churn
• burst_x_short_life, abrupt_decline → vol_churn
• sustained_engagement, exploration_score → engagement

1. CatBoost GPU — native categorical handling, TotalF1:Macro eval metric
2. LightGBM CPU — with target encoding for high-cardinality country_code
3. XGBoost GPU — sample weighting for class balance
4. Hierarchical 2-stage model:
   • Stage 1: churned vs not_churned (binary)
   • Stage 2: vol vs invol (binary, trained on churned subset)
   • Combined: P(vol) = P(churned) × P(vol|churned)

• Isotonic calibration per model before stacking
• Weighted average search across 4 models
• Per-class threshold tuning on OOF predictions (step=0.01)

• Burst pattern: vol_churn users generate heavily on day 0-1, then quit (median lifespan: 1.1 days!)
• High engagement → sudden drop: gen_trend strongly negative
• Japanese users: 60% vol_churn rate (cultural signal?)
• Explore feature, leave unsatisfied

• Prepaid cards: 2x higher invol rate (10.3% vs 4.9%)
• Indian users: 46% invol rate (payment infrastructure issues)
• Turkey/Brazil: 32%+ invol rate
• card_declined: 97% of all failed transactions
• CVC not_provided → risky card signal

• Subscription plan = useless (exactly 50/25/25 across all 4 plans)
• MI max = 0.035 → weak individual signals, model finds patterns in combinations

1. Day 2 re-engagement trigger — push notification/email if day01_ratio > 0.8 and silent on day 2
2. Personal discounts when gen_trend < -0.3 (usage declining)
3. Feature education for users with frustration = limited/high-cost
4. Premium content for power users (gen_total > median)
5. Loyalty rewards on day 7 milestone

1. Smart retry logic for card_declined (97% of failures)
2. Pre-dunning notifications before card expiry
3. Prepaid card warning — suggest credit alternatives at signup
4. Grace period (3-7 days) before deactivation on failure
5. Country-specific: Alternative payment methods for India/Turkey/Brazil
6. Digital wallet promotion to reduce decline rate

SHAP TreeExplainer on CatBoost models (averaged across 5 folds).

• Global importance: Top features per class (shap_global.png)
• Per-user explanations: Top-5 SHAP values for each prediction
• Class-specific SHAP profiles: Different drivers for vol vs invol

Per-user probability scores available in submission_detailed.csv.

.
├── hacknu_pipeline_kaggle.ipynb    # Main pipeline notebook
├── submission.csv                  # Test predictions (user_id, churn_status)
├── submission_detailed.csv         # Predictions + probabilities + confidence
├── shap_global.png                 # SHAP global importance by class
├── business_insights.png           # Behavioral analysis visualizations
└── README.md

1. Upload dataset to Kaggle (or Google Drive for Colab)
2. Update DATA_ROOT path in cell 3
3. Select GPU accelerator (T4 / H100)
4. Run all cells

Runtime: ~15-20 min on H100 / ~30-40 min on T4

Dependencies:

catboost, lightgbm, xgboost, optuna, shap, scikit-learn, pandas, numpy

• ML: CatBoost, LightGBM, XGBoost, Scikit-learn
• Explainability: SHAP
• Infrastructure: Kaggle Notebooks (H100 GPU)
• Languages: Python 3.10+

A 3-class classifier combining gradient boosting models (CatBoost/LightGBM/XGBoost) with a hierarchical 2-stage binary classifier. Features engineered across 6 data tables with emphasis on temporal patterns (burst detection, daily retention) and payment risk signals. Isotonic calibration and per-class threshold tuning applied for final predictions. SHAP-based explainability provides per-user churn reason attribution, enabling targeted retention strategies distinguishing voluntary (engagement-based) from involuntary (payment-based) churn.