machine-learning

Repository for the machine learning semester project: a cold-start recommendation system (notebook training + MLflow), a Flask inference API, and a React UI.

Setup

python -m venv venv
# Windows: venv\Scripts\activate
pip install -r requirements.txt

Train or obtain user_id_map.pkl and item_id_map.pkl (produced by the notebook’s load_data), and an MLflow-exported PyTorch model directory (or use the model registry).

Environment variables (inference API)

The API loads config from the process environment. Optional: create a .env file in the project root (python-dotenv is used by inference_api.py).

Variable	Default	Description
USER_ID_MAP_PATH	user_id_map.pkl	Path to the pickled user id → index map
ITEM_ID_MAP_PATH	item_id_map.pkl	Path to the pickled item id → index map
MLFLOW_MODEL_URI	mlartifacts/.../model (see inference_api.py)	Local path to an MLflow model folder (contains MLmodel), or a URI such as models:/ModelName/Stage or runs:/<run_id>/model when using MLflow tracking/registry

If map files are missing or a local model directory is invalid, the API exits at startup with an error message.

Run the API

From the repository root (so pickle paths resolve):

python inference_api.py

Default listening port: 5001.

Run the frontend (development)

cd recommendation-frontend
npm install
npm start

The dev server proxies /api to http://localhost:5001 (setupProxy.js). The React app calls /api/... by default.

Optional: set REACT_APP_API_BASE (e.g. http://localhost:5001) if you need an absolute API origin instead of same-origin /api.