New mitigation framework reduces bias in classification results

We use computers to help us make (hopefully) fair decisions. The problem is that machine learning algorithms won't always make fair classifications if the data used to train them has embedded human biases (in practice, (often).

To alleviate this “garbage in, garbage out” situation, the research team presented a flexible framework to reduce bias in machine classification. Their research is intelligent computing.

Existing attempts to reduce classification bias are often hampered by their reliance on specific fairness metrics or predetermined bias conditions, the team said. The team framework avoids these two kinds of dependencies. Bias reduction can be assessed based on various fairness metrics and infers specific bias terms from the data.

The team evaluated the framework on seven datasets across 21 machine classifiers. Throughout the experiment, the bias of the classification results was significantly reduced and the classification accuracy was almost maintained, achieving the desired results under the trade-off between fairness and practicality.

This framework shares the setup of an adversarial bias reduction method considering a proposal/consideration scenario between Alice, e.g., a company, and Bob, e.g., a regulator. Alice sends Bob a proposal to develop a target classifier, such as a university matching system, using Bob's data.

Bob considers the proposal and aims to ensure that Alice's classification does not exhibit any substantial bias along sensitive aspects that he is trying to protect, such as the student's middle school transfer experience. The goal is to build a classifier that minimizes discrimination along protected dimensions while incurring only a small performance penalty in target classification.

Bias mitigation is achieved by identifying data attributes that are prone to bias and applying effective data transformations to records based on these attributes.

This involves evaluating the contribution of attributes to data separation, calculating distances between attributes, and using these distances to establish bias attribute mappings within the constructed bias hyperspace. This mapping infers bias terms, recognizes bias-prone attributes, and measures their bias concentration.

However, applying workflows to large datasets can cause problems due to scalability limitations and other reasons.

In future work, the research team is interested in extending the framework to directly balance between fairness and accuracy of classification, taking into account potential conflicts between the public and private sectors. I have. From a broader perspective, incorporating behavioral features into classification bias reduction and analyzing practical settings in the application of such frameworks are important directions.