Personalizing the Prediction: Interactive and Interpretable machine learning

Abstract

While many applications with machine learning provide enough utilities for users, they mostly target average of users. Although it might be acceptable in certain domains, there are domains such as health and medical-care where it is crucial to provide personalized service. In such cases, personalization of machine learning model usually does not depend on end users to make change to the system. As machine learning models are black-box, the only information that the users can acquire is input and output of certain decision made by the model. Thus, with no reason behind specific prediction provided by the system, users cannot understand how the system works and make amendments to the system. This shortcoming is directly related to users' credibility in the system. In this paper, we present an interface where the system provides users the reason behind the decision made by the machine learning model and users provide feedback to the model. Moreover, we present the principle behind the suggested interface and prototype that instantiates the suggested interface. Our interface's effectiveness is evaluated through users' surveys regarding two main attributes: (1) how well users understand the system and more importantly, (2) how it influences users to trust in the system.