README

Project Overview

This repository contains data and scripts for training supervised classifiers (e.g., neural networks) on labeled acoustic datasets. The project is organized into three subsets:

• Subset_A/
• Subset_B/
• Subset_C/

Each subset has the following folder structure:

• 1_Original_Data/

  • original_data_for_classification.csv: Main dataset containing labels and extracted features required for building a supervised classifier.
  • Labeled_Data_With_Features.xlsx: Source file used to generate original_data_for_classification.csv. This file should not be modified.

• 2_Clean_Data/

  • Labeled_Data_to_Review.xlsx: File used to manually review and filter the dataset.
  • cleaned_data_for_classification.csv: Cleaned version of the dataset. This is the dataset to be used for training the second set of models.

The repository includes a Jupyter notebook that was used to create both original_data_for_classification.csv and cleaned_data_for_classification.csv from the source files (Labeled_Data_With_Features.xlsx and Labeled_Data_to_Review.xlsx).

The training and evaluation pipeline implemented is contained within train_and_eval.ipynb. SVM models are tuned individually for both versions of every subset. Metrics reported include:

• F1 Score: Overall and per-fold.
• Accuracy: Overall and per-fold.
• Final Confusion Matrix
• Statistical Testing: A Wilcoxon signed-rank test on F1 scores across folds is used to determine if the improvement is statistically significant.
• Visual Summary: Boxplots compare overall F1-Score and Accuracy performance across both versions of all subsets.