E-Thesis 181 views 53 downloads
Explainable Artificial Intelligence across Domains: Refinement of SHAP and Practical Applications / VEERA KOVVURI
Swansea University Author: VEERA KOVVURI
-
PDF | E-Thesis – open access
Copyright: The Author, Veera Raghava Reddy Kovvuri, 2024 Distributed under the terms of a Creative Commons Attribution 4.0 License (CC BY 4.0).
Download (5.71MB)
DOI (Published version): 10.23889/SUThesis.67149
Abstract
Explainable Artificial Intelligence (XAI) has become a crucial area within AI, emphasizing the transparency and interpretability of complex models. In this context, this research meticulously examines diverse datasets from medical, financial, and socio-economic domains, applying existing XAI technique...
Published: |
Swansea University, Wales, UK
2024
|
---|---|
Institution: | Swansea University |
Degree level: | Doctoral |
Degree name: | Ph.D |
Supervisor: | Seisenberger, M. |
URI: | https://cronfa.swan.ac.uk/Record/cronfa67149 |
Tags: |
Add Tag
xai
1
|
Abstract: |
Explainable Artificial Intelligence (XAI) has become a crucial area within AI, emphasizing the transparency and interpretability of complex models. In this context, this research meticulously examines diverse datasets from medical, financial, and socio-economic domains, applying existing XAI techniques to enhance understanding and clarity of the results. This work makes a notable contribution to XAI by introducing the Controllable fActor Feature Attribution (CAFA) approach, a novel method that categorizes dataset features into ‘controllable’ and ‘uncontrollable’ groups. This categorization enables a more nuanced and actionable analysis of feature importance.Furthermore, the research proposes an extension to CAFA, the Uncertainty-based Controllable fActor Feature Attribution (UCAFA) method, which incorporates a Variational Autoencoder (VAE) to ensure that perturbations remain within the expected data distribution, thereby enhancing the reliability of feature attributions. The effectiveness and versatility of CAFA are showcased through its application in two distinct domains:Medical and socio-economic. In the medical domain, a case study is conducted on the efficacy of COVID-19 non-pharmaceutical control measures, providing valuable insights into the impact and effectiveness of different strategies employed to control the pandemic. Additionally, UCAFA is applied to the medical domain, demonstrating its ability to improve the reliability of feature attributions by considering uncertainty. The socio-economic domain is investigated by applying CAFA to several datasets, yielding insights into income prediction, credit risk assessment, and recidivism prediction. In the financial domain, the analysis focuses on global equity funds using established XAI methodologies, particularly the integration of the XGBoost model with Shapley values. This analysis provides critical insights into fund performance and diversification strategies across G10 countries. This thesis highlights the potential of CAFA and UCAFA as promising directions in the domain of XAI, setting the stage for advanced research and applications. |
---|---|
Item Description: |
A selection of content is redacted or is partially redacted from this thesis to protect sensitive and personal information. |
Keywords: |
Explainable Artificial Intelligence, Artificial Intelligence, Machine Learning |
College: |
Faculty of Science and Engineering |