Abstract
Increasing implementations of machine learning models are not surprising due to their contribution to guiding humans in their decisions. Achieving competitive performance in machine learning (ML) models increases the complexity of understanding their predictions. Since competitive ML models often use complex algorithms and architectures, such as deep neural networks, it becomes challenging to intuitively understand how they arrive at specific predictions. This risks the solidity of their accuracy and trust in their decisions. This promotes the research field called eXplainable Artificial Intelligence (XAI) to help understand the rationales behind ML models by explaining models' prediction behaviour. This dissertation makes two key contributions to XAI: accelerating the process of obtaining a true model-agnostic explanation and consolidating conflicting explanations into a method-agnostic consistent explanation.As a contribution towards finding a true explanation, the thesis provides a concise overview of the history of XAI, the current state-of-the-art techniques, and potential future challenges. A significant challenge is associated with the shortcoming of explaining only the model's decisions, rather than the real-world inference in the data when dependent features exist. The improper handling of dependent features has led to an underestimation of those features' actual contribution towards the model's predictions. To address this issue, a feature importance explanation method that deals with the grouping of features is presented.
Furthermore, to accelerate the computation of feature importance, this thesis introduces a heuristic greedy search approach for dynamically determining the order of feature coalitions. The proposed model explainer is evaluated on complex ensemble models, including Random Forest and Gradient Boost classifiers, using both real-world and synthetic datasets. The research outcomes successfully capture the true inferences within the underlying data by equitably attributing importance to the equally informative features.
As a contribution towards finding a consistent explanation, the thesis provides a brief of model multiplicity, explanation multiplicity, and the related challenges posed to the XAI community. Multiple explanation methods return conflicting explanations for a model. It leads to confusion and uncertainty regarding the underlying behaviour of the model and poses significant challenges in identifying the most appropriate explanation.
This thesis delves into the issues arising from inconsistent explanations derived from various model-agnostic feature importance explanation methods and investigates strategies to achieve a consistent explanation by integrating and harmonising multiple conflicting explanations of ML models. The research evaluates various state-of-the-art explainers on multiple almost equally accurate Rashomon models, trained on customised real-world datasets and analyses the disagreements between their explanations. Ultimately, the research culminates in the provision of a unified explanation that is consistent across multiple methods, validating models' prediction behaviours.
| Date of Award | 23 Nov 2023 |
|---|---|
| Original language | English |
| Supervisor | Minsi Chen (Main Supervisor), Pritesh Mistry (Co-Supervisor) & Richard Hill (Co-Supervisor) |