# notebooks

Jupyter notebooks for interactive model training and evaluation. These mirror the CLI training scripts but with inline visualisations for loss curves, LOPO results, and confusion matrices.

## Notebooks

| Notebook | Purpose |
|----------|---------|
| `mlp.ipynb` | MLP training with per-epoch loss/accuracy plots, LOPO evaluation, confusion matrix, and ROC curve generation |
| `xgboost.ipynb` | XGBoost training with feature importance visualisation, LOPO evaluation, per-person metrics table, and ROC curve generation |

## When to use

Use the notebooks for exploratory work and visualisation during development. For reproducible training runs with ClearML logging and checkpoint saving, use the CLI scripts (`python -m models.mlp.train` and `python -m models.xgboost.train`).

## Data dependency

Both notebooks call into `data_preparation.prepare_dataset` to load and split the dataset. Collected `.npz` files must be present under `data/collected_<participant>/`. See [`data_preparation/README.md`](../data_preparation/README.md) for details on the data format and collection protocol.

## Key results reproduced

The notebooks produce the same core evaluation results as the CLI scripts:

**Pooled split (70/15/15)**

| Model | Accuracy | F1 | ROC-AUC |
|-------|----------|-----|---------|
| XGBoost | 95.87% | 0.959 | 0.991 |
| MLP | 92.92% | 0.929 | 0.971 |

**LOPO (9 participants)**

| Model | LOPO AUC | Optimal threshold | F1 at optimal |
|-------|----------|-------------------|---------------|
| MLP | 0.862 | 0.228 | 0.858 |
| XGBoost | 0.870 | 0.280 | 0.855 |