III: Small: Collaborative Research: Towards Interpretable Machine Learning

点击翻译按钮获取中文摘要

This research project investigates the design and development of machine learning algorithms that make decisions that are interpretable by humans. As predictions of machine learning models are increasingly used in making decisions with critical consequences (e.g., in medicine or economics), it is important that decision makers understand the rationale behind these predictions. The project defines interpretable algorithms through three key properties; Simplicity: intuitively comprehensible by users who are not experts in machine learning, Verifiability: a clear relationship between input features and model output, and Actionability: For a given input and desired output, the user should be able to identify changes to the input features that transform the model prediction to the desired output. The project investigates how to design distance metrics supporting simplicity and verifiability, as well as algorithms to identify input changes to change outputs. The project will be evaluated in a medical context, addressing the problem of early detection of hospital patients at risk of sudden deterioration.This work builds on the well-understood k-Nearest-Neighbor classifier, which would inherently seem to provide simplicity and verifiability. The challenge is in high dimensions, e.g., when used for document classification; differences are spread across more dimensions than are humanly comprehensible. The project uses novel dimensionality reduction approaches to create dissimilarity metrics that are interpretable and accurate. Visualization techniques to present this data will be explored, including techniques supporting more complex classification approaches such as ensembles. The project investigates novel methods for delivering actionability in machine learning algorithms by identifying changes that can truly transform an entity's class membership - a problem that has recently been identified as surprisingly difficult. A secondary outcome will be improvements in classifier robustness, as small changes that change class membership are a good indication of non-robustness.