A User Study on the Effect of Aggregating Explanations for Interpreting Machine Learning Models

Recently, there is growing consensus of the critical need to have better techniques to explain machine learning models. However, many of the popular techniques are instance-level explanations, which ex- plain the model from the point of view of a single data point. While local explanations may be misleading, they are also not human-scale, as it is impossible for users to read explanations for how the model behaves on all of their data points. Our work-in-progress paper explores the effectiveness of providing instance-level explanations in aggregate, by demonstrating that such aggregated explanations have a significant impact on users' ability to detect biases in data. This is achieved by comparing meaningful subsets, such as differences between ground truth labels, predicted labels, and correct and incorrect predictions, which provide necessary navigation to explain machine learning models.