GSTN Predictive Binary Classification¶

Role Solo ML Engineer & Lead

Timeline Aug 2024 – Oct 2024 (45 Days)

Dataset Scale 900,000+ Records (21 Attributes)

Primary Stack Python, XGBoost, LightGBM, SHAP

Finalist Selection: Ranked among the top 17 finalist teams out of 200+ national participating teams as a single-member solo developer.

Architecture & ML Pipeline Flow¶

graph TD
    A["900,000+ Anonymized GST Records"] --> B["Data Integrity Validation (SHA256)"]
    B --> C["Pre-processing &amp; Imputation (Median / Winsorization)"]
    C --> D["Class Imbalance Remediation (RUS + scale_pos_weight)"]
    D --> E["Stratified 5-Fold Nested Cross-Validation"]
    E --> F["Ensemble Modeling (XGBoost + LightGBM)"]
    F --> G["Threshold Tuning for F1 Optimization"]
    G --> H["SHAP Feature Interpretability Analysis"]
    H --> I["Competition-Compliant Model Artifact"]

Executive Overview¶

Developed for the Goods and Services Tax Network (GSTN) AI/ML Hackathon organized by the Government of India, this project engineered a high-throughput, interpretable binary classification pipeline for GST financial tax analytics.

The challenge required building an accurate predictive model $F_\theta(X) \to Y_{\text{pred}}$ over 900,000 real-world records characterized by severe class imbalance (91% majority / 9% minority) and extreme feature skewness, while adhering to strict zero-data-leakage compliance protocols.

Technical Challenges & Architectural Solutions¶

1. Severe Class Imbalance (91% / 9%)¶

Challenge: Standard loss functions biased predictions toward the majority class, causing unacceptably low minority recall.
Solution: Evaluated Random Under-Sampling (RUS), SMOTE, and tuned gradient boosted scale_pos_weight parameters to systematically optimize the Precision-Recall trade-off, maximizing both F1 and Matthews Correlation Coefficient (MCC).

2. Extreme Missingness & Heavy-Tailed Skewness¶

Challenge: Multiple tax feature columns exhibited >50% missing values and extreme financial outliers.
Solution: Applied strict feature pruning thresholds, robust median imputation, and two-sided Winsorization to normalize distribution tails without sacrificing variance.

3. Data Leakage & Generalization Safeguards¶

Challenge: Risk of subtle data leakage across feature engineering and hyperparameter search.
Solution: Enforced strict nested cross-validation and pipeline encapsulation (scikit-learn Pipeline) ensuring preprocessing transformations were fitted exclusively on training splits.

Performance & Evaluation Metrics¶

Evaluation Metric	Cross-Validation Score	Test Partition Score	Objective
Accuracy	97.6%	~97.8%	Global classification correctness
F1 Score	0.884	~0.891	Harmonic mean of precision and recall
MCC (Matthews Correlation)	0.875	~0.880	Balanced quality metric for imbalanced classes
ROC-AUC	0.988	~0.990	Separability threshold performance

Diagnostic Visualizations¶

Precision-Recall Curve

Confusion Matrix

Verified Accreditation¶

GSTN AI/ML National Hackathon Finalist • Awarded by Goods & Services Tax Network (GSTN)

Source Repository¶

GitHub Repository — mrxsierra/gstn_dsp_pbc: Complete reproduction scripts, cross-validation benches, and documentation.