Classification is a machine learning procedure where a class label is predicted from input data. Feature selection is a fundamental preprocessing technique, which aims to eliminate irrelevant features and reduce search space dimension. Feature selection is more complex if search space is substantial, sparse, or feature interaction is excessive. Selection of a relevant feature subset is vital, so as to decrease dimensionality, enhance accuracy, and reduce computational resource. Feature selection metrics with contextual accuracy or complexity abound. This thesis addresses current challenges in choosing a straightforward, accurate, generally applicable feature selection method, because machine learning is ubiquitous, automated, and expertise is frequently lacking. Practical challenges in rank filter feature selection and classification methods are addressed, like bias and threshold determination, to retain accuracy and stability. Pragmatic concerns are identified; useful requisite advice is given for domain experts, coders, and analysts. Techniques and metrics can then be more universally and confidently applied in many and varied contexts. This thesis formulates, details, and applies two novel hybrid feature selection filter classification metrics, U-Score and Uf-Score, based on merging uncorrelated heuristics. Designed to drastically reduce data set dimension, they true-rank relevant features with stability and accuracy. Uf-Score specifically eliminates redundant features of high inter-correlation, to further reduce dimension. Also, in organisational initiative trials, data classification issues required novel feature selection approaches; human-centred algorithm design (HCAD) developed within real-life contexts. Design ethnographic human-centred algorithm design (DE HCAD) is conceived, frame-worked and honed, thereby modernising dimension reduction and bias perspectives for feature selection. U-Score and Uf-Score may be used independently or alongside an original participatory design ethnographic (PDE HCAD) framework. It is established and validated for bottom-up and/or top-down organisational or systems change; it reimagines bias with broad reflective interpretation. A unique, complete schema is achieved via a user-friendly feature selection decision flowchart. This thesis verifies innovative, comprehensive hybrid dimensionality reduction feature selection methods with simple, stable, and accurate metrics, affording clear-cut interpretation. They extend seamlessly to human-centred algorithm design processes, appropriate for complex, changing, or uncertain contexts, and can result in practical, beneficial, and more forward-looking solutions.
| Date of Award | 25 Mar 2026 |
|---|
| Original language | English |
|---|
| Supervisor | Richard Hill (Main Supervisor) & Qiang Xu (Co-Supervisor) |
|---|