Document Type
Thesis
Degree Name
Master of Applied Computing
Department
Physics and Computer Science
Program Name/Specialization
Applied Computing
Faculty/School
Faculty of Science
First Advisor
Yang Liu
Advisor Role
Thesis Supervisor
Abstract
As machine learning has advanced significantly over the past decade, predictive models have achieved substantial success across various domains. However, they often lack transparency in their decision-making processes. This opacity presents risks, particularly in sectors like healthcare and finance, where transparency is crucial. To address these challenges, explainable artificial intelligence (XAI) has risen in importance. Techniques such as LIME (Local Interpretable Model-agnostic Explanations) have been developed, utilizing simpler alternative models like linear regression or tree-based models to explain decisions for specific instances.
This thesis initially concentrates on the application of LIME to three global annual datasets, specifically within the relatively uncharted territory of explaining data with temporal changes. Through the utilization of LIME, we can effectively pinpoint the crucial features that influence predictions. Additionally, the incorporation of FLAML facilitates the exploration of various predictive methods, thereby enabling us to attain robust performance. Furthermore, we have executed a range of data preprocessing techniques to improve the quality of our predictions. To validate LIME's effectiveness, Individual Conditional Explanation (ICE) plots are employed, grounding the explanations in observable data trends. This approach not only improves the understanding of LIME's application in rendering complex datasets interpretable but also highlights the significance of feature importance within predictive models.
Building on this foundational understanding, the thesis then proposes a novel counterfactual generation method. Counterfactual explanations are a type of XAI to help users understand model decisions by posing 'what-if'' scenarios. These explanations focus on describing how slight changes in the input data could lead to different output results, effectively answering the question, "What needs to change in the input for a different decision to be made by the model?'' Characterized by the goal of simultaneously achieving validity, proximity and sparsity, our proposal employs simulated annealing optimization and post-pruning techniques to produce counterfactuals that are both effective and comprehensible. Specifically, LIME is adopted in the optimization process by prioritizing features and reducing the search space of candidate counterfactuals.
By advancing these methods, the thesis contributes to the field of explainable AI, providing strategies that not only help interpret model decisions but also empower users to influence future outcomes based on these insights.
Recommended Citation
Nakano, Shou and Liu, Yang, "Advancing Machine Learning Interpretability: LIME Applications to Temporal Data and A Novel Method of Counterfactual Generation" (2024). Theses and Dissertations (Comprehensive). 2667.
https://scholars.wlu.ca/etd/2667
Convocation Year
2024
Convocation Season
Fall