Presented by: Thies Lindenthal (University of Cambridge)
A rapidly expanding universe of technology-focused startups is trying to change and improve the way real estate markets operate. The undisputed predictive power of machine learning (ML) models often plays a crucial role in the ‘disruption’ of traditional processes. However, an accountability gap prevails: How do the models arrive at their predictions? Do they do what we hope they do – or are corners cut?
Training ML models is a software development process at heart. We follow best practices from software engineering and define a system testing framework to verify that the trained ML models perform as intended. Illustratively, we augment two real-estate related image classifiers with a system testing stage based on local interpretable model-agnostic explanation (LIME) techniques. Analysing the classifications sheds light on some of the factors that determine the behaviour of the systems. We show that cross-validation is simply not good enough when operating in regulated environments.