Title: A Methodological Framework on Accountability in Societal Applications of Machine Learning
Abstract:
Machine learning (ML) techniques are increasingly used in dealing with the growing complexity of data analysis encountered in various societal industries, including healthcare, e.g. prediction and prognosis of chronic diseases, drug discovery, financial risk management, fraud detection, manufacturing, and forensics. However, there are also serious concerns that the decisions based on ML techniques could lead to undesirable consequences for individuals and society. The current lack of sufficient trust in ML applications mainly stems from the fact that ML algorithms are so intricate that it is virtually impossible and highly impractical for relevant stakeholders to fully understand how they process the data and thereby yield results which are of significance to individuals and society. The accountability of decisions based on ML techniques is necessary to build and maintain public trust in ML applications. In the context of societal applications of ML, accountability means the responsibility that the relevant stakeholders who are affected by the decisions based on ML techniques should be provided with adequate explanations regarding the credibility of these decisions as well as their possible consequences. In this talk, I will regard ML as a modelling activity that is laden with the values and interests of various stakeholders involved in ML applications. I will argue that accountability in the context of ML applications cannot be properly dealt with without considering the kinds of value judgments made by specialists in designing ML systems and evaluating their performance in accordance with the interests and values of relevant stakeholders. I will thus propose a value-based methodological framework within which accountability in the context of societal applications of ML can be interpreted and assessed.
