manhngvu/dna_noc

DNA Mixture Analysis - Unknown Contributor Detection (Optimized v2.0)

📊 Dataset Overview

This dataset contains optimized machine learning models and data for detecting unknown contributors in DNA mixture samples.

Key Metrics

• F1 Score: 0.7135 ✅ (optimized, +5.66% improvement)
• Recall: 0.8600 (86% detection rate)
• Precision: 0.6262 (balanced for forensics)
• Samples: 500 total (408 RD14 + 92 RD12 DNA kits)
• Features: 260 domain-specific DNA metrics
• Evaluation: 5-fold stratified cross-validation

🗂️ Directory Structure

├── data/
│   ├── combined_enhanced_features.csv (500 × 260)
│   ├── rd14_enhanced_features.csv (408 × 260)
│   └── rd12_enhanced_features.csv (92 × 260)
├── results/
│   ├── cross_validation_results.csv
│   ├── performance_summary.csv
│   ├── performance_by_fold.csv
│   └── data_statistics.json
├── reports/
│   ├── EXECUTIVE_SUMMARY.md
│   ├── TECHNICAL_REPORT.md
│   ├── OPTIMIZATION_REPORT.md
│   └── DEPLOYMENT_GUIDE.md
└── code/
    ├── inference_pipeline.py
    ├── optimization_params.json
    └── dataset_summary.json

🎯 Performance (Optimized)

Model	F1 Score	Precision	Recall
Ensemble	0.7135	0.6262	0.8600
XGBoost	0.7482	0.6548	0.9000
CatBoost	0.6324	0.5595	0.7667

Optimization Strategy

• Weights: XGBoost 0.70, CatBoost 0.30
• Threshold: 0.42 (optimized for forensic recall)
• Improvement: +0.0382 F1 (+5.66%) from baseline

📚 Documentation

Quick Start

Start with EXECUTIVE_SUMMARY.md for:

• Problem overview
• Key results
• Improvement over baseline
• Production deployment status

Technical Details

See TECHNICAL_REPORT.md for:

• Complete methodology
• Feature engineering (260 domain-specific features)
• Cross-validation results
• Error analysis
• Future work

Implementation

See DEPLOYMENT_GUIDE.md for:

• Installation instructions
• Python API usage
• Integration examples
• Configuration options

🔧 How to Use

Load Data in Python

import pandas as pd

# Load combined features
df = pd.read_csv('data/combined_enhanced_features.csv')

# 500 samples × 260 features
print(df.shape)  # (500, 260)

# Access CV results
cv_results = pd.read_csv('results/cross_validation_results.csv')

Use Inference Pipeline

from code.inference_pipeline import DNAMixtureAnalyzer

analyzer = DNAMixtureAnalyzer(model_dir='.')
prediction = analyzer.predict(sample_features)

print(f"Unknown present: {prediction['unknown_present']}")
print(f"Confidence: {prediction['confidence']:.1%}")

📊 Dataset Characteristics

Class Distribution

• No Unknown Contributors: 470 samples (94%)
• Has Unknown Contributors: 30 samples (6%)
• Imbalance Ratio: 15.7:1

DNA Kits

• RD14: 408 samples (81.6%)
• RD12: 92 samples (18.4%)

Features (260 total)

• Peak heights (count, max, sum, mean, std)
• Peak ratio (peak₁/peak₂)
• Allele balance (min/max peaks)
• Homozygosity indicator (binary)
• Peak consistency (coefficient of variation)
• Applied across 20+ STR markers

🎓 Citation

@dataset{manhngvu_dna_noc_v2,
  title={DNA Mixture Analysis: Unknown Contributor Detection (Optimized v2.0)},
  author={Nguyen, Manh},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/manhngvu/dna_noc}
}

🚀 Use Cases

• Forensic DNA Analysis: Detect unknown contributors in mixture samples
• Case Screening: Prioritize cases for further investigation
• Research: Benchmark ensemble methods on imbalanced DNA data
• Reproduction: Complete methodology and code for scientific reproducibility

✅ Quality Assurance

• ✅ 500 complete samples with 260 features each
• ✅ No missing values after preprocessing
• ✅ Stratified 5-fold cross-validation
• ✅ Hyperparameter optimization (Optuna)
• ✅ No data leakage in evaluation
• ✅ Reproducible with fixed random seeds

📞 Support

For questions about:

• Data: See data_statistics.json
• Results: See results/ directory
• Implementation: See code/ directory
• Methodology: See reports/ directory

---

Version: 2.0 (Optimized)
Updated: 2026-05-08
Status: Production-Ready ✅