How svarshneysjsu/predicting-internet-use-severity-in-child-with-actigraphy works
Project Type & Purpose
• Type: Machine-learning pipeline / data-science workflow for a Kaggle competition.
• Purpose: Ingest raw health-and-activity data, preprocess and engineer features, train multiple models (TabNet, LSTM, MLP, ensemble), evaluate, and submit predictions.
File Structure
• Top-level artifacts only:
• README.md – project overview, data sources, models, metrics
• cmi_prediction_using_VRandTabNet.ipynb – Jupyter notebook combining code for data prep, modeling, evaluation.
• Team8_Project_Report.pdf – final project report.
• No separate “src” or “scripts” directories, so treat the notebook as the single orchestration unit.
README for Key Components
• Data sources: Healthy Brain Network (tabular and time-series accelerometer).
• Stages: Preprocessing → EDA → Feature Engineering → Modeling → Evaluation → Submission.
• Models: TabNet, custom TabNet wrapper, LSTM variants, MLP, Voting Regressor ensemble.
• Tools / Frameworks: Python, Jupyter, PyTorch-TabNet, TensorFlow/Keras or PyTorch for LSTM, LightGBM / XGBoost / CatBoost, Borderline-SMOTE.
1. Data Sources
• Raw Tabular CSVs (demographics, usage, sleep)
• Raw Time-Series Accelerometer Logs
2. Data Ingestion & Storage
• File system or DataFrame loader
3. Data Preprocessing
• Missing-value handling, scaling, Borderline-SMOTE
4. Exploratory Data Analysis (EDA)
• Visualization modules (matplotlib / seaborn)
5. Feature Engineering
• Tabular features
• Time-series feature extraction (windowing, summary stats)
• Combined feature set
6. Modeling
• Branch A: TabNet Classifier / Custom Wrapper
• Branch B: Sequential models (LSTM, Siamese LSTM)
• Branch C: MLP on tabular features
7. Ensemble Module
• Voting Regressor combining LightGBM, XGBoost, CatBoost, TabNet
8. Evaluation & Metrics
• QWK calculation (train vs. validation)
• Kaggle submission file generator
9. Output & Reporting
• Submission CSV
• Visualizations saved to disk
• Team8_Project_Report.pdf
b. Relationships & Data Flow
Data Sources → Ingestion → Preprocessing → (EDA / Feature Eng.) → Modeling branches → Ensemble → Evaluation → Submission & Report
c. Architectural Patterns & Principles
• Pipeline pattern: sequential stages.
• Branch-and-merge pattern for parallel model training and ensemble.
• Modularity: separate preprocessing, feature engineering, modeling.
d. Key Technologies & Frameworks
• Python / Jupyter Notebook (orchestration)
• Pandas / NumPy for data prep
• Borderline-SMOTE (imbalanced-learn)
• PyTorch-TabNet for TabNet models
• TensorFlow/Keras or PyTorch for LSTM
• scikit-learn for MLP and voting regressor
• LightGBM, XGBoost, CatBoost for ensemble
Detailed Elements
• Data source origins (HBN dataset) with notes on data types.
• Preprocessing sub-steps (imputation, scaling, SMOTE).
• Parallel model branches and their frameworks.
• Ensemble combining logic and metrics computation.
• Final outputs: Kaggle submission and project report.