abhayra12/StudentLife-Phenotyping
End-to-end behavioral prediction system using digital phenotyping. PyTorch Transformer (MAE 1.18) + Autoencoder anomaly detection. Docker-ready, FastAPI service.
What's novel
End-to-end behavioral prediction system using digital phenotyping. PyTorch Transformer (MAE 1.18) + Autoencoder anomaly detection. Docker-ready, FastAPI service.
Code Analysis
9 files read · 3 roundsTrains and serves ML models (gradient boosting ensembles, PyTorch Transformer, LSTM, Autoencoder) that predict student stress levels and physical activity from passive smartphone sensor data (accelerometer, audio, WiFi, phone-lock) using the StudentLife dataset, exposed via a FastAPI service.
Strengths
Genuine end-to-end ML pipeline with real algorithmic depth: Optuna HPO, soft-voting ensembles, PyTorch Transformer with positional encoding, weekend-normalized autoencoder anomaly detection, and a well-structured data alignment layer that resamples 9 heterogeneous sensor streams onto a common hourly grid. MLflow experiment tracking is consistently integrated across all deep learning models.
Weaknesses
Significant code duplication (create_sequences repeated in LSTM and Transformer, feature engineering logic duplicated across stress_prediction.py and sota_stress_prediction.py), the API's anomaly interpretation thresholds are arbitrary and disconnected from the training-time percentile-based thresholds, and the test suite is minimal (only 2 test files covering data cleaning, no tests for models, API, or feature engineering).
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
71
Contributors
1
Files
108
Active weeks
6
Repository
Language
Python
Stars
1
Forks
0
License
MIT