ECEN 758 Data Mining and Analysis - Fall 2025

Team Members:

• Banghyon Lee (Joseph)
• Fahimeh Orvati Nia
• Nandhini Valiveti
• Sushma Perati

This project implements comprehensive environmental sound classification using the ESC-50 dataset. We explore traditional machine learning methods, deep learning approaches, and transfer learning techniques to classify 50 different environmental sound categories.

Live Website: https://fahimehorvatinia.github.io/Group20-Data-Mining-Project/

The website includes interactive visualizations, model performance comparisons, and detailed results analysis.

ESC-50 Dataset:

• 2000 audio recordings (5 seconds each)
• 50 environmental sound classes
• 40 examples per class
• Pre-arranged into 5 folds for cross-validation

The dataset includes diverse environmental sounds across 5 categories:

Sample audio waveforms from different sound classes

Mel spectrograms showing frequency content over time for different sound classes

3D PCA projection of features showing class separability (explained variance on each axis)

• MFCC (Mel-frequency cepstral coefficients): 20 features capturing spectral envelope
• Chroma features: 12 features representing pitch class information
• Mel spectrogram: 128 mel-scaled frequency bands
• Spectral contrast: 7 features capturing spectral peaks and valleys
• Tonnetz: 6 features representing tonal characteristics

Spectrograms showing time-frequency representations of different sound classes

• YAMNet embeddings: 2048 features per audio file (mean + std pooling of 1024-dim frame embeddings)
  • Pre-trained on AudioSet (2 million audio clips)
  • Captures high-level audio representations

• Raw features
• Normalized features (StandardScaler)
• Selected features (Top 100 via Mutual Information)

• Naive Bayes Classifier
• 1D Convolutional Neural Network (CNN)

• Random Forest
• Gradient Boosting
• XGBoost
• Transfer Learning with YAMNet + Random Forest
• MLP on YAMNet embeddings
• Self-Supervised Learning (Pseudo-Labeling)

• Support Vector Machine (SVM) - Linear kernel
• Support Vector Machine (SVM) - RBF kernel
• Multi-Layer Perceptron (MLP)

• Logistic Regression
• K-Nearest Neighbors (KNN)

All models are evaluated using:

• Accuracy
• Precision (macro-averaged)
• Recall (macro-averaged)
• F1-Score (macro-averaged)
• Confusion Matrices
• Learning Curves
• Feature Importance (where applicable)

.
├── ESC_50_Complete_Project_All_Methods.ipynb  # Main project notebook
├── Project_Report.tex                          # IEEE format report
├── results.html                                # Results visualization website
├── data_out/                                   # Feature files (CSV)
│   ├── esc50_features_raw.csv
│   ├── esc50_features_normalized_corrected.csv
│   └── esc50_features_selected.csv
├── ESC-50_data/                               # Dataset (audio files excluded from repo)
│   └── meta/                                   # Metadata only
└── README.md                                  # This file

pip install numpy pandas scikit-learn matplotlib seaborn tensorflow tensorflow-hub librosa xgboost jupyter

1. Clone this repository:

git clone https://github.com/fahimehorvatinia/Group20-Data-Mining-Project.git
cd Group20-Data-Mining-Project

2. Download the ESC-50 dataset from here and extract to ESC-50_data/

3. Open and run ESC_50_Complete_Project_All_Methods.ipynb in Jupyter Notebook

4. View results in results.html (generated after running the notebook)

• Unified data splitting ensuring all classes in train/val/test
• No data leakage between splits
• Comprehensive model comparison
• Transfer learning with YAMNet
• Self-supervised learning exploration
• GPU/CPU compatibility

• Transfer Learning Dominance: YAMNet embeddings achieve 83-84% test accuracy vs. 16-47% for traditional features, representing a 35+ percentage point improvement
• Best Traditional Model: Random Forest achieves 46.8% accuracy on normalized features
• Feature Preprocessing: Normalization improves average accuracy from 28% to 31% for traditional classifiers
• Self-Supervised Learning: Pseudo-labeling provides 0.3-0.4% improvements over baseline YAMNet+RF

See the Results Website for detailed performance comparisons across all models and feature sets.

The project report is available in Project_Report.tex (IEEE conference format).

This project is for educational purposes as part of ECEN 758 Data Mining and Analysis course.

• ESC-50 dataset creators
• TensorFlow Hub for YAMNet model
• scikit-learn, TensorFlow, and other open-source libraries