Personalized News Recommendation System (Python)

A complete, end-to-end, modular recommendation system for news articles. It simulates or ingests data, preprocesses and engineers features, trains both classical and deep-learning recommenders, evaluates them offline, simulates A/B testing, and serves real-time recommendations via FastAPI.

Highlights

• Data simulation: news metadata and user interactions (clicks, time, likes)
• Preprocessing & EDA: cleaning, normalization, notebook walkthrough
• Feature engineering: TF-IDF (content), user profiles from history
• Classical recommenders: item-based CF (cosine), truncated-SVD MF
• Deep recommender: simple two-tower (user/item id embeddings) in PyTorch
• Offline evaluation: Precision@K, Recall@K, NDCG, MAP
• A/B testing simulator: online-style evaluation harness
• Deployment: FastAPI REST service + minimal web UI

Project Structure

.
├─ requirements.txt
├─ README.md
├─ notebooks/
│  └─ 01_eda.ipynb
├─ src/
│  ├─ __init__.py
│  ├─ data/
│  │  └─ simulate.py
│  ├─ preprocess.py
│  ├─ features.py
│  ├─ models/
│  │  ├─ classical.py
│  │  └─ deep.py
│  ├─ eval/
│  │  └─ metrics.py
│  ├─ abtest/
│  │  └─ simulator.py
│  └─ pipeline.py
├─ api/
│  └─ main.py
└─ frontend/
   └─ index.html

Quickstart

1. Create a virtual environment and install dependencies:

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

2. Run the training pipeline to generate artifacts:

python -m src.pipeline

This simulates data, preprocesses, builds features, trains recommenders, evaluates them, and saves artifacts to artifacts/.

3. Start the API server:

uvicorn api.main:app --reload

4. Open the minimal UI:

• Navigate to http://127.0.0.1:8000 to see recommendations for a random user.

Workflow (Text Flow)

Ingest/Simulate data → Clean & EDA → Feature engineering → Train (Classical + Deep) → Offline evaluation → A/B Simulation → Package artifacts → Serve via API → Optional frontend

Notebooks

• notebooks/01_eda.ipynb explains the dataset, behavior patterns, and feature distributions.

Notes

• Swap in real ingestion instead of simulation as needed.
• Baselines and deep model are simple and readable; extend as desired.
• Consider advanced embeddings (e.g., transformer text encoders) for richer content.