Teaching Meaningful Explanations

The adoption of machine learning in high-stakes applicatio ns such as healthcare and law has lagged in part because predictions are not accomp anied by explanations comprehensible to the domain user, who often holds ult imate responsibility for decisions and outcomes. In this paper, we propose an appr oach to generate such explanations in which training data is augmented to inc lude, in addition to features and labels, explanations elicited from domain use rs. A joint model is then learned to produce both labels and explanations from the inp ut features. This simple idea ensures that explanations are tailored to the compl exity expectations and domain knowledge of the consumer. Evaluation spans multipl e modeling techniques on a simple game dataset, an image dataset, and a chemi cal odor dataset, showing that our approach is generalizable across domains a nd algorithms. Results demonstrate that meaningful explanations can be reli ably taught to machine learning algorithms, and in some cases, improve modeling ac curacy.