Project README: ID2223 - Political Fact Checker

Overview

This project is the culmination of the course ID2223 - Scalable Machine Learning and Deep Learning. Our team has developed a political fact-checking tool leveraging an LSTM model trained on dynamic data. The project comprises four key components: backfilling pipeline, EDA and feature engineering pipeline, transformation and training pipeline, and deployment into Hugging Face Spaces.

Technologies Used

• Hopsworks: For the feature store.
• Modal: For function scheduling and deployment.
• Hugging Face Spaces and Gradio: For UI deployment.

Project Components

1. Backfilling Pipeline

• Data Source: Historical data was sourced from PolitiFact, scraping 600 data points.
• Process: The data was uploaded to the Hopsworks feature store.

2. EDA and Feature Engineering Pipeline

• Scheduling: A scheduler was deployed on Modal to retrieve all data points from the page daily.
• Data Storage: The retrieved data was uploaded to the Hopsworks feature store.

3. Transformation and Training Pipeline

• Data Transformation: This included label standardization and encoding, text tokenization and indexing, and converting the data into PyTorch tensors.
• Model Training: We used GloVe embeddings with 2 LSTM layers. Post-training, the model was uploaded to the Hopsworks model registry.

4. Deployment

• UI Interface: A simple UI was created on Hugging Face Spaces.

Potential Improvements

• Implement logic to avoid duplicate data points in the feature store.
• Increase the historical data size to improve model training and performance.

Evaluation

• The model showed inconsistent performance, indicated by large fluctuations in F1 scores, likely due to overfitting.
• Future improvements could focus on expanding the data set and refining the model to enhance accuracy.