BiasField: Interactive Bias Probing of Machine Learning Datasets

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Machine learning datasets can contain biases, where specific attributes are unfairly linked, impacting model performance and fairness.
• Current bias detection methods often overlook implicit biases tied to dataset-specific attributes like facial expressions or attire, focusing instead on explicit biases related to predefined categories (e.g., gender, ethnicity).

Purpose of the Study:

• To introduce BiasField, an interactive tool designed for the detection, analysis, and mitigation of bias in machine learning datasets.
• To address the limitations of existing methods by enabling the adaptive detection of both explicit and implicit biases using dataset-specific attributes.

Main Methods:

• BiasField automatically extracts dataset-specific attributes for adaptive bias detection.
• It utilizes a plant-growth metaphor for visualizing biases and analyzing their relationships.
• Bias mitigation is achieved through targeted generative data augmentation.

Main Results:

• The tool successfully detects, analyzes, and mitigates complex biases in datasets.
• User, case, and expert studies validated BiasField's effectiveness.

Conclusions:

• BiasField offers a comprehensive closed-loop workflow for managing bias in machine learning.
• It enhances fairness and model performance by addressing both explicit and implicit biases.