Multi-Modal Policy Success & Impact Prediction

📌 Overview

This project implements a Multi-Modal, Multi-Output Deep Learning Model designed to predict the success probability and impact score of government or organizational policies. By leveraging both the policy's written text and relevant socio-economic data, the model provides a comprehensive prediction.

The system also includes a Flask-based Web Application for user interaction, allowing policymakers and researchers to test policy inputs and receive real-time predictions through a simple interface.

⚙️ Methodology

1. Text Processing
2. Model: BERT (bert-base-uncased)

3. Converts policy text into a 768-dim embedding → reduced to 128-dim.

4. Numeric Feature Processing

5. Model: Multi-Layer Perceptron (MLP)
6. Input: 6 numeric features.
7. Architecture: Linear(6→64) → ReLU → Linear(64→32).

8. Output: 32-dim numeric embedding.

9. Fusion Layer

10. Concatenates text embedding (128-dim) + numeric embedding (32-dim).

11. Produces a shared 64-dim representation.

12. Outputs

13. Success Probability → Linear(64→1) + Sigmoid (0.0–1.0)

14. Impact Score → Linear(64→1) (-1.0–+1.0)

15. Explainability

16. LIME Tabular → Explains numeric features.
17. LIME Text → Highlights influential words in text.

📊 Features

🚀 How It Works

1. User inputs policy text + numeric features via Flask UI.
2. Model processes each modality separately (BERT for text, MLP for numeric).
3. A fusion layer merges embeddings.
4. Outputs:
5. Success Probability (0–1)
6. Impact Score (-1 to +1)

🖥️ Web Application (Flask UI)

The project includes a Flask web app for real-time predictions.
The interface allows users to input data and visualize model outputs.

📷 Place your Flask app screenshot at:

🛠️ Tech Stack

• Python 3.9+
• Flask
• PyTorch
• Hugging Face Transformers (BERT)
• scikit-learn
• LIME (Explainability)
• Pandas / NumPy

📂 Project Structure

├── data/               # Input datasets
├── models/             # Model architecture & training scripts
├── notebooks/          # Jupyter notebooks for experiments
├── utils/              # Preprocessing & helper functions
├── static/images/      # Web app static assets (UI screenshot here)
├── app.py              # Flask app entry point
├── templates/          # HTML templates for Flask
├── Report.docx         # Project documentation
└── README.md           # Project overview

📖 Example Input/Output

Input Example (JSON):

{
  "policy_text": "Increase renewable energy subsidies to reduce carbon emissions.",
  "population": 170000000,
  "gdp_per_capita": 4000,
  "education_index": 0.65,
  "past_success_rate": 0.45,
  "past_impact_score": 0.3,
  "avg_sentiment": 0.7
}

Output Example (JSON):

{
  "success_probability": 0.78,
  "impact_score": 0.52
}

📌 Future Improvements

• Integration with SHAP for more robust interpretability.
• Expanding features with additional socio-economic indicators.
• Deploying as a REST API for real-time evaluation.

👨‍💻 Authors

• Isha Das — 📧 isha@idtechsolutionsbd.com
• Md Jisan Ahmed — 📧 jisan3325@gmail.com