A Field Experiment on Group Bias in Machine Learning Classification

How to cite

Asensio, Omar Isaac; Ha, Sooji; Marchetto, Daniel. 2018. "A Field Experiment on Group Bias in Machine Learning Classification." AEA RCT Registry. AEARCTR-0003425. https://doi.org/10.1257/rct.3425-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Omar Isaac Asensio* , Sooji Ha , Daniel Marchetto

* Primary investigator

Registry ID
AEARCTR-0003425
DOI
https://doi.org/10.1257/rct.3425-1.0
First registered
Last updated
Status
In development
Trial dates
2018-11-01 – 2019-03-31
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/3425

Abstract

Many algorithms that conceal hidden biases are being used in business and government to make important decisions in society. For example, historical data used to train machine learning (ML) algorithms are known to unintentionally discriminate against minority groups (e.g. by gender, race, etc.) depending on the nature of the observational data. Researchers typically have little to no control over the sample characteristics of the training data and often rely on technical solutions after-the-fact. Field experiments can be an effective alternative to rigorously test and potentially mitigate the effects of algorithmic bias in a number of applications. In this study, we implement a randomized controlled trial (RCT) to detect and mitigate possible bias in the context of electric vehicle (EV) usage. Expert and non-expert participants will complete a classification task using reviews from a popular electric vehicle mobile app. The data collected will be used for training a classification algorithm using neural network-based language models. In addition, policy preferences will be collected during the intervention to determine if the review classification tasks influence participant beliefs as they relate to electric vehicle infrastructure.

Experimental design

We will deploy a survey questionnaire to build a training dataset for machine learning classification tasks. Human raters will be asked to read short text related to electric vehicle experiences and will be asked to classify each review as either positive or negative and identify one or more topic categories that describe the text from a list of predefined choices. All reviews that include potentially sensitive information such as cell phone number or email have been redacted using the terms [phone number] or [email address]. Each participant will classify a total of 20 randomly selected reviews from a set of 127,257 publicly available electric vehicle charging station reviews, as provided by a popular EV charging station locator app. After categorizing the reviews, participants will be asked for short demographic information. Pre- and post- task policy preferences will be collected for analysis. No personally identifiable information will be collected or shared.

Design details

Randomization method
The randomization algorithm will holdout approximately 2,000 randomly selected reviews per campaign without replacement in order to calculate inter-rater reliability score. The remaining reviews will be randomly sampled from a universe of approximately 125,000 reviews using a computer generated randomization procedure.
Randomization unit
We are randomizing reviews at the individual level. For the holdout sample for inter-rater reliability, we are randomizing over a cluster of 3 individuals per review in order to activate wisdom of crowds and tie-breaker in case of disagreement.
Clusters
We have approximately 127,257 station reviews. After separating the holdout sample, we will randomly assign the remaining 125,257 reviews into 1000 clusters of 15 reviews. Each respondent will be assigned 20 reviews, 15 of which come from the clusters of 15 and 5 from the holdout for scientific reliability measures.
Observations
We will have 1000 U.S. general population participants completing approximately 17,000 reviews (15 x 1000 + 2000 holdout) to build the training set, which is greater than our minimum 10% sampling target (12,725 reviews) for the training set. We will then repeat the sampling procedure to 1000 subject matter experts for 17,000 additional observations, for a total of 34,000 observations in the final training set for classification.
Arms
1000 participants will be drawn from the U.S. general population and 1000 participants will be drawn from a crowd of subject matter experts for a total of 2000 participants in the full study.
Primary outcomes
The key outcomes include: (i) topic classifications for contents of electric vehicle station reviews and (ii) policy preferences on support for EV programs.

Cited by

No citation data available for this registration yet.

Keywords