This application implements a recommendation system using a two-tower neural network architecture, integrated with Feast for feature management. It generates synthetic datasets for users, items, and interactions, trains a model to produce user and item embeddings, and provides personalized item recommendations.

• Data Generation: Creates synthetic datasets and images under generation directory
• Feature Store: Utilizes Feast to manage and serve features, with configurations defined in feature_repo/.
• Model Training: Trains a two-tower model (UserTower and ItemTower) to generate embeddings, implemented in models/.
• Filtering: Applies rule-based filtering (availability, demographic, history, and contextual) to refine recommendations (models/filtering.py).

The application includes two main data generation components:

Generates synthetic e-commerce data including:

• Users: Creates user profiles with IDs, signup dates, and category preferences
• Items: Generates product data with categories, prices, ratings, and descriptions based on Amazon-like structure
• Interactions: Simulates user-item interactions (views, purchases, ratings) with preference-based bias

The generated data is saved in parquet format under feature_repo/data/:

• recommendation_users.parquet
• recommendation_items.parquet
• recommendation_interactions.parquet

# Generate synthetic data
python generation/dataset_gen_amazon.py --n_users 1000 --n_items 5000 --n_interactions 20000

Generates product images using Stable Diffusion:

• Uses the RunwayML Stable Diffusion v1.5 model
• Creates images based on product descriptions
• Supports both CPU and CUDA-enabled GPU processing
• Saves generated images in PNG format under generation/data/generated_images/

To generate the datasets, run:

# Generate product images (requires PyTorch and diffusers)
python generation/generate_images.py