This repository compares the performance of static and constructive neural networks. Results from the corresponding writeup (V1) can be reproduced as follows:

1. Run data_prep.ipynb to process data/features_by_participant.xlsx.
2. Run train_test.ipynb. The training cell must be re-run for each dataset, which is done by changing the variable dataset at the top of the cell.

Note that 5-run k-fold cross-validation can take a while to run.

In addition to packages from the standard library, you'll need:

sklearn, torch, matplotlib, seaborn, numpy, pandas, ipympl, openpyxl.

All other python files are helpers:

• constr_casc.py defines the constructive cascade network architecture (Casper).
• static_nets.py defines the static network architecture (both for logistic regression and MLP training).
• performance_metrics.py contains helpers for tracking the performance of each model.
• data_prep.ipynb processes the original dataset.

To track results across sessions, performance metrics are saved to .csv files under results/. The results included with the current repository reflect those presented in the writeup linked above. Each time a model is run/evaluated, results for the corresponding dataset and model ID will be overwritten in the appropriate CSV file, unless recompute_model_perfs is removed from the end of the training cell.

For references, see the writeup.