Trust AI & Data Quality Pipeline

Overview

Trust AI & Data Quality Pipeline is a comprehensive framework designed to enhance data quality (DQ) and integrate explainable AI (XAI) methodologies. It addresses challenges in real-world data processing and model reliability by establishing automated workflows that optimize data preprocessing, boost interpretability, and enable robust decision-making.

Key Features

1. Data Quality Enhancement

   • Implements both parametric and non-parametric imputation methods such as Optimized Fuzzy C-Means (OFCM) and other ML/DL-based approaches.
   • Evaluates imputation effectiveness using metrics like Differential Entropy, Canonical Relative Entropy, and JS-Divergence.

2. Explainable AI (XAI) Framework

   • Active Learning: Adapts dynamically to user requirements while ensuring efficient labeling through knowledge distillation.
   • SHAP Interpretability: Provides feature-level importance explanations.
   • Counterfactual Explanations: Offers actionable insights by generating alternative scenarios that would lead to different predictions.

3. Pipeline Automation

   • Combines the data quality module (EDASH) and XAI components to automate end-to-end workflows, ensuring reliable and interpretable AI solutions.

---

Architecture

Data Quality Module: EDASH

1. Imputation Algorithms

   • Optimized Fuzzy C-Means (OFCM): Incorporates weighting factors to optimize cluster accuracy and minimize overfitting.
   • Other methods include Expectation-Maximization and MissForest.

2. Evaluation Metrics

   • Differential Entropy: Measures data distribution before and after imputation.
   • JS-Divergence: Quantifies similarities between datasets.
   • NRMSE (Normalized Root Mean Square Error): Assesses reconstruction quality.

---

Trust AI: Explainable AI (XAI) Modules

1. Active Learning

   • Query Strategy: Includes uncertainty-based sampling and margin sampling to optimize labeled data acquisition.
   • Human Annotation: Ensures high-confidence labeling for critical datasets.

2. SHAP (Shapley Additive Explanations)

   • Visualizes feature importance to highlight model decision logic.
   • Enhances trust by presenting interpretable insights to users.

3. Counterfactual Explanations

   • Provides "what-if" scenarios to facilitate decision-making by adjusting critical feature values.

---

Case Study: Gas Sensor Dataset

• Achieved an F1 score of 0.9928 in predictive tasks.
• Imputation tests demonstrated superior reconstruction accuracy using OFCM, outperforming baseline methods by >20%.
• Active Learning reduced labeling costs by >30% while maintaining high model accuracy.

---

Demo

Demo Video

Watch the Demo Video

Website Preview

---

Implementation Details

Technology Stack

• Frontend: React.js with Tailwind CSS for responsive UI.
• Backend: Django with PostgreSQL, containerized using Docker.
• Imputation Module: Python with NumPy, SciPy, and custom ML algorithms.

---

Running the System

Frontend:

cd frontend
npm install
npm run dev

• Accessible at: http://localhost:5173

Backend & Database:

cd backend
docker compose up --build

• Django API runs on: http://127.0.0.1:8000

---

References

1. EDASH Repository: GitHub - EDASH
2. Presentation Materials: Poster