FairQueryRepair-ML

Machine Learning Guided Fair Query Repair with Cross-Dataset Adaptation.

---

Project Overview

Modern database queries may satisfy all SQL predicates while still violating real-world fairness requirements. This project investigates how unfair queries can be automatically repaired by modifying query predicates so that the resulting query satisfies fairness constraints.

The work builds upon a fairness-aware query repair framework and extends it with machine learning techniques for repair prediction, confidence estimation, and ML-guided search.

---

Problem Statement

Given:

• A user query
• A fairness constraint (e.g., Statistical Parity Difference (SPD) ≤ 0.20)

Find:

• A minimally modified query (repair) that satisfies the fairness constraint while preserving query validity.

---

Original Framework

The original framework:

1. Generates unfair queries
2. Measures fairness using Statistical Parity Difference (SPD)
3. Searches for valid repairs
4. Returns the top-k repaired queries

Pipeline:

Unfair Query → Fairness Evaluation → Repair Search → Fair Query

---

Proposed Extensions

1. Machine Learning-Based Repair Prediction

A Random Forest model is trained on historical query-repair pairs to predict repair actions.

Three progressively richer datasets were created:

• V1: Query thresholds and repair targets
• V2: Added fairness statistics
• V3: Added repair-distance information

2. ML-Guided Search

Pure ML predictions are used as a starting point for local neighborhood exploration.

Pipeline:

ML Prediction → Neighbor Generation → Fairness Validation → Best Repair

3. Cross-Dataset Adaptation

The framework was adapted from the Adult Income dataset to the COMPAS recidivism dataset to evaluate cross-domain applicability.

---

Experimental Results

Repair Prediction

Method	Success Rate
Pure ML Repair	28%
ML-Guided Search	87%

Feature Importance

Fairness Improvement

Dataset	SPD Before	SPD After
Adult	0.248	0.175
COMPAS	0.191	0.113

---

Datasets

Adult Income Dataset

Used to generate:

• training_repairs_v1.csv
• training_repairs_v2.csv
• training_repairs_v3.csv

COMPAS Dataset

Used for cross-dataset adaptation:

• compas_repairs_v2.csv

---

Repository Structure

datasets/
├── training_repairs_v1.csv
├── training_repairs_v2.csv
├── training_repairs_v3.csv
└── compas_repairs_v2.csv

figures/

notebooks/
├── final_experiments.ipynb
└── cross_dataset.ipynb

results/

paper/

---

Reproducibility

Main experiments can be reproduced using:

• notebooks/final_experiments.ipynb
• notebooks/cross_dataset.ipynb

---

Future Work

• Cross-domain transfer learning
• Additional fairness metrics
• Multi-objective query repair
• Adaptive repair search strategies