RecommendationSystem

Deployed Website available at https://trumandaniels.dev/recommendation-system/

Portfolio recommendation-system demo built around real user-app install interactions from the Myket Android Application Install Dataset.

The product story is:

Given the apps a user has installed, recommend the next apps they are likely to install.

The frontend compares two recommendation strategies:

• Popularity: globally most-installed apps in the Myket interaction sample.
• Item-Item CF: apps often installed by users who also installed the selected user's apps.

The Myket data is a user-app-time interaction graph, so it also gives this project a working PyTorch Geometric training path for a lightweight LightGCN-style item-item scorer.

Local Data

The source dataset is kept under ignored data/ paths:

data/myket-android-application-market-dataset/
data/myket.db

To rebuild the SQLite database and generated frontend data:

python3 injest_data.py --replace-db
python3 scripts/export_demo_data.py

PyTorch Geometric Scoring

Install the pinned model dependencies:

sfw .venv/bin/pip install -r requirements.txt

Train a bounded CPU-friendly graph model:

.venv/bin/python train_models.py train-pyg-item-item \
  --max-users 2000 \
  --max-apps 2000 \
  --max-edges 100000 \
  --epochs 5 \
  --output-path .runtime/models/pyg-item-item.pt

Score from an existing user's install history:

.venv/bin/python train_models.py recommend-pyg-item-item \
  --model-path .runtime/models/pyg-item-item.pt \
  --user-id -1152799605 \
  --limit 10

Frontend

cd frontend
npm run dev

For a production check:

cd frontend
npm run build