Manipulating and Measuring Model Interpretability

要旨

With machine learning models being increasingly used to aid decision making even in high-stakes domains, there has been a growing interest in developing interpretable models. Although many supposedly interpretable models have been proposed, there have been relatively few experimental studies investigating whether these models achieve their intended effects, such as making people more closely follow a model's predictions when it is beneficial for them to do so or enabling them to detect when a model has made a mistake. We present a sequence of pre-registered experiments (N = 3,800) in which we showed participants functionally identical models that varied only in two factors commonly thought to make machine learning models more or less interpretable: the number of features and the transparency of the model (i.e., whether the model internals are clear or black box). Predictably, participants who saw a clear model with few features could better simulate the model's predictions. However, we did not find that participants more closely followed its predictions. Furthermore, showing participants a clear model meant that they were *less* able to detect and correct for the model's sizable mistakes, seemingly due to information overload. These counterintuitive findings emphasize the importance of testing over intuition when developing interpretable models.

著者
Forough Poursabzi-Sangdeh
Microsoft Research, NYC, New York, United States
Daniel G. Goldstein
Microsoft Research, New York, New York, United States
Jake M. Hofman
Microsoft Research, NYC, New York, United States
Jennifer Wortman Vaughan
Microsoft Research, New York, New York, United States
Hanna Wallach
Microsoft Research, New York City, New York, United States
DOI

10.1145/3411764.3445315

論文URL

https://doi.org/10.1145/3411764.3445315

動画

会議: CHI 2021

The ACM CHI Conference on Human Factors in Computing Systems (https://chi2021.acm.org/)

セッション: Human, ML & AI

[A] Paper Room 14, 2021-05-10 17:00:00~2021-05-10 19:00:00 / [B] Paper Room 14, 2021-05-11 01:00:00~2021-05-11 03:00:00 / [C] Paper Room 14, 2021-05-11 09:00:00~2021-05-11 11:00:00
Paper Room 14
13 件の発表
2021-05-10 17:00:00
2021-05-10 19:00:00
日本語まとめ
読み込み中…