# Common-Sense Bias Discovery and Mitigation for Classification Tasks

**Authors:** Miao Zhang, Zee fryer, Ben Colman, Ali Shahriyari, Gaurav Bharaj  
**arXiv:** [2401.13213](https://arxiv.org/abs/2401.13213)  
**Date:** 2024-01-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2401.13213v1/

## Key points

- Machine learning model bias can arise from dataset composition
- Sensitive features correlated to the learning target can disturb the model's decision rule
- Existing de-biasing methods focus on capturing image features in the latent space, but this is insufficient to understand all dataset feature correlations
- The authors propose a framework called Common-Sense Bias Discovery (CSBD) that extracts feature clusters based on image descriptions
- CSBD captures both subtle and coarse features of images and utilizes human-in-the-loop examination for analysis
- The analyzed features and correlations are human-interpretable
- Downstream model bias can be mitigated by adjusting image sampling weights without requiring sensitive group label supervision
- Experiments on benchmark image datasets show that CSBD discovers novel biases and outperforms state-of-the-art unsupervised bias mitigation methods
- CSBD leverages image descriptions to extract feature clusters and discover biases in datasets
- CSBD provides insights into dataset feature correlations and offers effective mitigation strategies for model bias without relying on sensitive group labels.

## Summary

The paper titled "Common-Sense Bias Discovery and Mitigation for Classification Tasks" addresses the issue of machine learning model bias that can arise from dataset composition. The authors highlight that sensitive features correlated to the learning target can disturb the model's decision rule, leading to performance differences along these features. Existing de-biasing methods have focused on capturing prominent and delicate image features traceable in the model's latent space. However, they argue that using the latent space alone is insufficient to understand all dataset feature correlations. To address this limitation, the authors propose a framework called Common-Sense Bias Discovery (CSBD) that extracts feature clusters in a dataset based on image descriptions. This approach allows them to capture both subtle and coarse features of the images. They formulate the feature co-occurrence pattern and measure correlation, utilizing a human-in-the-loop for examination. Importantly, the analyzed features and correlations are human-interpretable. By exposing sensitive correlations in a dataset, the authors demonstrate that downstream model bias can be mitigated by adjusting image sampling weights without requiring sensitive group label supervision. They conduct experiments on two benchmark image datasets for multiple classification tasks and show that their method discovers novel biases. Furthermore, they compare their intervention with state-of-the-art unsupervised bias mitigation methods and find that it outperforms them. Overall, this paper presents an innovative framework, CSBD, which leverages image descriptions to extract feature clusters and discover biases in datasets. The proposed method not only provides insights into dataset feature correlations but also offers effective mitigation strategies for model bias without relying on sensitive group labels.