Statistical Machine Learning Coursework

Overview

This coursework project applies supervised and unsupervised machine learning techniques to three distinct datasets:

• Blood Cell Classification using K-Nearest Neighbours (KNN)
• Loan Risk Prediction using Random Forest
• Gaia Star Clustering using UMAP + HDBSCAN The aim is to provide a brief overview of a variety of ML techniques to a range of different dataset.

---

The Report

Report.pdf is the full written report submitted for marking. It achieved a 100% grade.

---

Requirements

• Python 3.13.11
• See requirements.txt for full list of dependencies

Install dependencies

pip install -r requirements.txt

---

Dataset

Three datasets are required to run this project:

• blood_cell_anomaly_detection.csv — Blood cell features and cell type labels
• loan_risk_prediction_dataset.csv — Loan applicant features and approval labels
• gaia-dr2-rave-35.csv — Gaia DR2 star catalogue with astrometric features

---

Usage

Run the full pipeline:

python src/pipeline.py

---

1. Blood Cell Classification (KNN)

K-Nearest Neighbours is applied to classify blood cell types. Features are standardised using StandardScaler before training. K values from 1 to 24 are evaluated to find the optimal number of neighbours.

Best K: 15

KNN Accuracy vs K

Confusion Matrix (K=15)

---

2. Loan Risk Prediction (Random Forest)

A Random Forest classifier with 200 estimators and max depth of 8 is trained to predict loan approval. Categorical variables are one-hot encoded prior to training. Out-of-bag scoring is enabled to give an additional estimate of generalisation performance.

Single Decision Tree from Forest

Feature Importance

---

3. Gaia Star Clustering (UMAP + HDBSCAN)

100,000 stars are sampled from the Gaia DR2 catalogue. Features are standardised and reduced to 2 dimensions using UMAP for visualisation. HDBSCAN is then applied to identify clusters, with noise points (label = -1) removed from the final plot.

HDBSCAN Parameters:

• min_cluster_size = 300
• min_samples = 10
• cluster_selection_method = 'eom'

UMAP Projection

HDBSCAN Clustering (Noise Removed)