ML Model & XAI Pipeline

This repository contains an end-to-end Machine Learning pipeline for tabular data, featuring robust preprocessing, multi-model training, and advanced Explainable AI (XAI) reports.

Core Features

• Categorical Support: Automatically detects and encodes categorical features using OneHotEncoder.
• Data Leakage Prevention: Automatically identifies and drops source columns used to derive target labels to ensure realistic evaluation.
• Multi-Class Robustness: Handles binary and multi-class classification with weighted performance metrics (Precision, Recall, F1) and One-vs-Rest ROC-AUC.
• Interactive Explanations: Generates HTML-based ELI5 reports for global feature weights and local prediction debugging.
• Dynamic Configuration: Fully parameterized scripts support custom input paths, target columns, and tuning parameters via CLI.

---

Installation

Create a virtual environment and install the required dependencies:

# Create venv
python -m venv venv

# Activate venv (Windows)
.\venv\Scripts\Activate.ps1

# Install dependencies
pip install pandas scikit-learn matplotlib eli5 numpy

---

Usage Workflow

Step 1: Data Preprocessing

Clean and scale your raw dataset. This script handles missing values (median/most-frequent imputation), standardizes numeric features, and One-Hot encodes categorical data.

python preprocess.py --input device1_top_20_features.csv --output preprocessed_data.csv --label label

• Default Input: device1_top_20_features.csv
• Output: Scaled and encoded .csv ready for training.

Step 2: Model Training & XAI

Train Decision Tree, Random Forest, and Neural Network (MLP) models.

python xai.py --input preprocessed_data.csv --source_col HH_L5_pcc

• Arguments:
  • --input: Path to the file from Step 1.
  • --source_col: The feature used to derive targets (if any). This column will be dropped to prevent Data Leakage.
  • --num_examples: Number of local prediction explanations to generate (default: 5).

---

Outputs

1. Performance Metrics: Detailed console output including Accuracy, Weighted F1-Score, and Weighted OVR ROC-AUC.
2. feature_importances.png: A high-resolution comparison plot of the top 12 features for each model.
3. eli5_explanation_<Model>.html: Interactive HTML reports containing:
   • Global Summary: Feature weights and decision logic.
   • Local Explanations: Breakdown of individual predictions for specific samples.

Performance Benchmark (device1 dataset)

Model	Accuracy	F1 (Weighted)	ROC-AUC
Decision Tree	~0.81	~0.81	~0.94
Random Forest	~0.82	~0.82	~0.97
Neural Network	~0.83	~0.83	~0.97

---

Notes

• Make sure to use the provided venv for all executions to ensure dependency compatibility.
• For multi-class datasets, metrics are calculated using Weighted Averages to account for class imbalance.