Explainable machine learning and deep learning

Abstract

The ability of machine learning to improve research, processes and products is significant. Nonetheless, one major challenge hindering its adoption is that computers do not provide explanations for the predictions made in general. This project aims to solve this problem by focusing on methods for making machine learning models and decisions made interpretable to humans.

In this report, the main concepts related to interpretability are stated first. Next, both global and local model-agnostic methods are explored and implemented to interpret specific models or certain predictions made by the models. Each method implemented is elaborated in detail, including how it functions, its advantages and the negative effects.

Two datasets, the bike sharing dataset and the cervical cancer dataset, are used as examples to explain and analyze different methods used in this project on regression and classification levels.