Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Identifying decision boundaries to explain black-box machine learning predictions
Thesis (PhD)--Stellenbosch University, 2026.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
Stellenbosch : Stellenbosch University
2026
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613744364781568 |
|---|---|
| access_status_str | Open Access |
| author | Rowan, Adriaan Izak |
| author2 | Lubbe, Sugnet |
| author_browse | Lubbe, Sugnet Rowan, Adriaan Izak |
| author_facet | Lubbe, Sugnet Rowan, Adriaan Izak |
| author_sort | Rowan, Adriaan Izak |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (PhD)--Stellenbosch University, 2026. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/135750 |
| institution | Stellenbosch University (South Africa) |
| language | English |
| last_indexed | 2026-06-10T12:41:00.939Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2026 |
| publishDateRange | 2026 |
| publishDateSort | 2026 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/135750 Identifying decision boundaries to explain black-box machine learning predictions Rowan, Adriaan Izak Lubbe, Sugnet Stellenbosch University. Faculty of Economics and Management Sciences. Dept. of Statistical and Actuarial Science. Machine learning -- Interpretation Decision making -- Statistical methods Artificial intelligence -- Evaluation Computational intelligence UCTD Thesis (PhD)--Stellenbosch University, 2026. Rowan, A. I. 2026. Identifying decision boundaries to explain black-box machine learning predictions. Unpublished doctoral dissertation. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/0b725a04-f095-45ea-bb0c-da12f2881a79 ENGLISH SUMMARY: The increasing deployment of machine learning models in high-stakes decision making has intensified the demand for interpretability methods that enable both practitioners and stakeholders to understand, trust and challenge model predictions. Existing approaches are limited; many are computationally intensive, restricted to either local or global explanations, or lack model-agnostic applicability. This study developed a new post hoc explanation methodology that integrates principal component and canonical variate analysis biplot techniques with counterfactual reasoning to provide visual interpretations that are plausible, feasible and actionable. The framework uses biplots to generate two-dimensional visualisations that approximate multi-dimensional distances between data points. An inverse projection function enables variable values to be extracted from the two-dimensional surface while preserving the geometric relationships of the original high-dimensional data. This allows decision boundaries to be approximated in reduced space and the nearest counterfactuals identified with respect to those boundaries. By mapping observed data points to their nearest counterfactuals, the method reveals not only which variables drive classification outcomes, but also the relative change required in each variable to alter those outcomes. Decomposing these counterfactual shifts provides local interpretability at the individual level and, when aggregated across the dataset, yields a global approximation of variable importance. This offers a computationally efficient alternative to Shapley-based approaches while retaining model-agnostic applicability. Counterfactual identification also allows the exact contribution of each variable to a prediction to be quantified, answering the fundamental interpretability question: “What minimal change in the input would have led to a different outcome?” Furthermore, the method links interpretability to robustness: Models with counterfactuals lying close to the decision boundary are more sensitive to small perturbations, while those with more distant counterfactuals exhibit greater stability. The uniqueness of this research lies in bridging geometric data visualisation with modern explainable AI (artificial intelligence) techniques. By re-purposing statistical tools traditionally used for dimension reduction, the study introduces an interpretable, visual and computationally efficient approach to understanding machine learning models. This contributes a novel perspective to the field of explainable AI and demonstrates its practical value in actuarial and risk-modelling contexts, where transparency, stability and accountability are paramount. AFRIKAANSE OPSOMMING: Die toenemende gebruik van masjienleermodelle vir belangrike besluitneming in die finansiele en mediese industriee het die behoefte aan interpreteerbaarheidsmetodes sterk laat toeneem. Praktisyns en belanghebbendes wil nie net weet wat ’n model voorspel nie, maar ook hoekom. Baie bestaande metodes het egter beperkings: sommige is berekeningsintensief, ander fokus net op plaaslike of globale verklarings, en baie werk slegs vir spesifieke modeltipes en is dus nie model-onafhanklik nie. Hierdie studie stel ’n nuwe post hoc-verklaringsmetodologie voor wat hoofkomponent- en kanoniese-variansie-analise- bistippingstegnieke kombineer met teenfeitlike redenering. Die doel is om visuele verklarings te verskaf wat geloofwaardig, uitvoerbaar en prakties bruikbaar is. Die raamwerk gebruik bistippings om tweedimensionele voorstellings te skep wat die multidimensionele afstande tussen datapunte benader. ’n Omgekeerde projeksiefunksie maak dit moontlik om veranderlike-waardes vanaf die tweedimensionele oppervlak terug te bereken, terwyl die meetkundige struktuur van die oorspronklike hoedimensionele data behoue bly. Dit stel die navorser in staat om besluitgrense in ’n laerdimensionele ruimte te benader en die naaste teenfeitlikes relatief tot daardie grense te identifiseer. Deur elke datapunt met sy naaste teenfeitlike te vergelyk, wys die metode watter veranderlikes ’n klassifikasie-uitkoms dryf en hoeveel elke veranderlike sou moes verander om die uitkoms te beinvloed. Die ontleding van hierdie teenfeitlike verskuiwings gee plaaslike interpreteerbaarheid op individuele vlak. Wanneer die resultate oor die hele datastel saamgevoeg word, verskaf dit ook ’n globale aanduiding van veranderlike-belangrikheid. Dit bied ’n meer berekenings doeltreffende alternatief tot Shapley-gebaseerde benaderings, terwyl dit steeds model-onafhanklik bly. Die identifisering van teenfeitlikes maak dit ook moontlik om die presiese bydrae van elke veranderlike tot ’n voorspelling te meet. Dit beantwoord die kernvraag van interpreteerbaarheid: “Watter minimale verandering in die inset sou tot ’n ander uitkoms gelei het?” Die metode toon verder ’n duidelike verband tussen interpreteerbaarheid en robuustheid: Modelle waarvan teenfeitlikes baie naby aan die besluitgrens le, is meer sensitief vir klein versteurings, terwyl modelle met verder gelee teenfeitlikes meer stabiel is. Die kernbydrae van hierdie navorsing is dat dit meetkundige datavisualisering met moderne verklaarbare KI-tegnieke kombineer. Deur statistiese gereedskap vir dimensievermindering op ’n nuwe manier te gebruik, bied die studie ’n interpreteerbare, visuele en berekenings doeltreffende manier om masjienleermodelle beter te verstaan. Dit wys hoe verklaarbare KI praktiese waarde kan toevoeg in aktuariele en risikomodelleringsomgewings, waar deursigtigheid, stabiliteit en aanspreeklikheid noodsaaklik is. Doctoral 2026-04-09T09:01:13Z 2026-04-09T09:01:13Z 2026-03 Thesis https://scholar.sun.ac.za/handle/10019.1/135750 en Stellenbosch University 218 pages : illustrations, includes annexures application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Machine learning -- Interpretation Decision making -- Statistical methods Artificial intelligence -- Evaluation Computational intelligence UCTD Rowan, Adriaan Izak Identifying decision boundaries to explain black-box machine learning predictions |
| title | Identifying decision boundaries to explain black-box machine learning predictions |
| title_full | Identifying decision boundaries to explain black-box machine learning predictions |
| title_fullStr | Identifying decision boundaries to explain black-box machine learning predictions |
| title_full_unstemmed | Identifying decision boundaries to explain black-box machine learning predictions |
| title_short | Identifying decision boundaries to explain black-box machine learning predictions |
| title_sort | identifying decision boundaries to explain black box machine learning predictions |
| topic | Machine learning -- Interpretation Decision making -- Statistical methods Artificial intelligence -- Evaluation Computational intelligence UCTD |
| url | https://scholar.sun.ac.za/handle/10019.1/135750 |
| work_keys_str_mv | AT rowanadriaanizak identifyingdecisionboundariestoexplainblackboxmachinelearningpredictions |