Extract: Explainable transparent control of bias in embeddings

This work addresses the challenge that knowledge-graph embeddings risked leaking protected attributes (e.g., gender, age, occupation) from behavioral data, raising privacy and fairness concerns.

Abstract: We developed EXTRACT, a suite of transparent methods for controlling bias in knowledge graph embeddings. The approach applies Canonical Correlation Analysis to pinpoint bias-leakage sources, decomposes embeddings into private-attribute vectors via linear-system solving, and integrates four transparent mitigation methods.

Key Contributions:

• EXTRACT suite for bias control in knowledge graph embeddings
• Canonical Correlation Analysis for bias-leakage detection
• Linear-system solving for embedding decomposition
• Four transparent mitigation methods
• Balance between model utility and privacy protection

Methodology:

• Engineered vector-based knowledge base construction
• Applied CCA to pinpoint bias-leakage sources
• Decomposed embeddings into private-attribute vectors
• Integrated transparent mitigation methods to strip unwanted signals

Results: Demonstrated robust recommending performance alongside bias control, highlighting the trade-off between accuracy and privacy.

Applications: Adapted bias-mitigation techniques to financial-entity embeddings, ensuring fair, regulatory-compliant representations for risk models and portfolio construction.

Recommended citation: Z. Guo, Z. Xu, M. Lewis, and N. Cristianini, “Extract: Explainable transparent control of bias in embeddings,” in AEQUITAS 2023: AEQUITAS 2023 First AEQUITAS Workshop on Fairness and Bias in AI co-located with ECAI 2023, 2023.