Kelvyn Bladen
Machine learning methods are powerful analytical tools used across all scientific disciplines and many other fields of investigation for prediction and inference from diverse data sources. Despite their broad applicability, machine learning methods are often highly complex and difficult to interpret. Developing a greater understanding of which variables most influence a response is essential for increasing the interpretability of these models and supporting informed decision-making. This research focuses on improving how we evaluate the importance of these variables. One common approach is to shuffle the values of a variable and see how much the model accuracy gets worse. Another approach removes a variable entirely to see how much it mattered. The first paper shows how these two machine learning ideas provide similar information to well-understood classical statistical techniques. This helps make machine learning measurements easier to interpret, by connecting them to tools we already understand. The second paper develops new mathematical results that explain why these connections occur, especially when variables are correlated with each other. In real data, variables are often correlated, which can make it harder to tell which ones are truly important. This work provides a way to account for correlation and offers clear mathematical equations showing how different importance methods are affected by this. Finally, the third paper introduces a new method called CLIQUE that explains predictions at the level of individual data points. Instead of describing what variables matter on average, it shows which variables are important for a specific observation while avoiding misleading results caused by interacting variables. Overall, this research helps make machine learning methods more transparent and trustworthy by connecting them to classical statistics and creating better tools for under standing both global patterns and individual predictions.