Practical Machine Learning – Feature Analysis & Multi-Label Classification

Assignments for CCE3503 – Practical Machine Learning, including feature selection techniques and multi-label neural network classification, supervised by Dr. Trevor Spiteri at the University of Malta.

---

Repository Contents

.
├── assignment1/
│   ├── feature_analysis.ipynb       # Notebook for feature selection and evaluation
│   └── ...                          # CSV dataset assumed to be downloaded via VLE
│
├── assignment2/
│   ├── multi-label-classification.ipynb   # Full pipeline for multi-label classification
│   └── ...                                # Yeast dataset (not included)
└── README.md

---

Assignment 1 – Feature Analysis (Communities & Crime Dataset)

Objective

Perform data cleaning and dimensionality reduction using:

• Filter methods (correlation heatmap + thresholding)
• Wrapper methods: Sequential Forward/Backward Selection (SFS, SBS)
• Projection: Principal Component Analysis (PCA)

Each technique was evaluated by training an MLPRegressor on a target variable: assaultPerPop.

Highlights

• Missing data imputed using mean strategy.
• Normalisation applied using StandardScaler.
• Train/test split: 80/20

Evaluation

MSE (Mean Squared Error) was used to compare:

• Filter
• SFS
• SBS
• PCA

Final results showed that PCA and SFS achieved lowest MSE, with PCA reducing feature count significantly.

---

Assignment 2 – Multi-Label Classification (Yeast Dataset)

Objective

Train and evaluate neural networks for multi-label classification using:

• Problem transformation (Binary Relevance vs Classifier Chains)
• Algorithm adaptation with hyperparameter optimisation (HPO)

Dataset

• Yeast gene expression dataset (2417 samples, 103 features, 14 classes)
• Preprocessing and label binarisation performed within notebook

Approaches

• Binary Relevance (BR): independent classifiers per label
• Classifier Chains (CC): models label dependencies
• Adapted Neural Network: trained with OneVsRestClassifier and MLPClassifier using HPO

Hyperparameters Tuned

• Hidden layer sizes
• Learning rate
• Batch size

HPO performed using GridSearchCV with 5-fold cross-validation.

Evaluation Metrics

• Hamming Loss
• Exact Match Ratio (EMR)
• F1 Score (Macro and Micro)

Observations

• Classifier Chains outperformed Binary Relevance in Exact Match Ratio.
• Adapted neural network showed best balance across all metrics.
• EMR values were expectedly low due to high label cardinality.

---

How to Run

1. Clone the repository:

git clone https://github.com/GrahamPellegrini/cce3503.git

2. Launch Jupyter Lab or VSCode and open the notebooks:

jupyter lab assignment1/feature_analysis.ipynb
jupyter lab assignment2/multi-label-classification.ipynb

3. Download the datasets via VLE and place them in the correct folders.

Communities & Crime CSV for Assignment 1
Yeast Dataset for Assignment 2

4. Run all notebook cells sequentially. Make sure all dependencies are installed:

pip install numpy pandas matplotlib seaborn scikit-learn scikit-multilearn

---

Author

Graham Pellegrini
B.Eng. (Hons) Computer Engineering
University of Malta
GitHub: @GrahamPellegrini