Efficiently Local Explanation for Black-box Machine Learning Model (Liang Lan et al.)

A big concern of applying machine learning models is the “black-box” nature of machine learning. One of the popular methods for machine learning model interpretation is to provide local explanation for each prediction. Recently proposed Shapley value provides a unified way to explain any complex classification model locally. The idea is based on fundamental concepts from coalitional game theory that compute individual contribution of each player to an overall reward. When apply Shapley value for interpreting machine learning model, each feature is a “player” in a coalitional game and the prediction is the overall reward. Despite the theoretical advantages of Shapley value for computing individual feature contribution, computing Shapley value requires an exponential time complexity that prevents it for practical use.

The objective of this project is to study efficient approximation of complex black-box machine learning models and to provide efficiently local explanation for their predictions. We will develop novel methods for approximating kernel methods and deep neural networks and therefore offering efficiently local explanation for these complex models.

Related Publications:


For further information on this research topic, please contact Dr. Liang Lan.