Diagnostics and Explainable Machine Learning Models

Despite the advances of Machine Learning, the models are still being considered black-boxes that are difficult to diagnose and explain. The model performance diagnostic measures are critical to the assessment of the model’s relevance, accuracy and robustness. Good models’ performance is the primary enabler of their successful deployment in real-life applications. However, even if the models perform well, it is not known why the models predict the way they do, that is, which input variables are responsible for the models’ predictions. The purpose of the research is two-prone: 1) to identify the relationship between the measures of model performance and recommend which measures should be used in the model production environment, and, 2) develop the methodology of explaining machine learning models in terms of the models’ inputs, as well as, other, potentially relevant, variables, not selected in the model.

Ningsheng Zhao
Faculty Supervisor: 
Jia Yuan Yu
Partner University: