BiasFairnessAccountabilitycs.HC
Measuring Social Biases of Crowd Workers using Counterfactual Queries
Bhavya Ghai, Q. Vera Liao, Yunfeng Zhang, Klaus Mueller
April 4, 2020
arXiv.org
17 citations
Bhavya Ghai, Q. Vera Liao, Yunfeng Zhang, Klaus Mueller
Crowdsourced datasets inherit worker biases: The paper demonstrates that social biases based on gender, race, and other factors don't just exist in ML models—they originate in the crowdsourced training data itself, making the data collection process a critical point where bias enters the pipeline.
Individual worker bias varies significantly: By measuring each crowd worker's inherent social biases separately, the research reveals that bias isn't uniform across the workforce, suggesting that some workers systematically introduce more bias than others into labeled datasets.
Counterfactual queries can measure bias: The paper proposes using counterfactual fairness methods (asking "what if" questions with changed demographic attributes) as a practical tool to quantify how biased individual workers are, providing a measurable metric rather than assumptions.
Bias measurement enables better dataset curation: The key practical application is that knowing each worker's bias level allows organizations to weight or filter worker responses when creating training datasets, potentially reducing the biased data that feeds into AI systems before models are even trained.
Cost-effective bias mitigation at the source: Rather than trying to debias datasets or models after the fact, this approach addresses bias early in the crowdsourcing stage, offering a more efficient way to improve dataset quality without necessarily increasing costs.
Social biases based on gender, race, etc. have been shown to pollute machine learning (ML) pipeline predominantly via biased training datasets. Crowdsourcing, a popular cost-effective measure to gather labeled training datasets, is not immune to the inherent social biases of crowd workers. To ensure such social biases aren't passed onto the curated datasets, it's important to know how biased each crowd worker is. In this work, we propose a new method based on counterfactual fairness to quantify the degree of inherent social bias in each crowd worker. This extra information can be leveraged together with individual worker responses to curate a less biased dataset.
Accepted at the Workshop on Fair and Responsible AI at ACM CHI 2020
Citation data from Semantic Scholar