Abstract
Knowledge Base (KB) embeddings are susceptible to societal biases present in their training data. This paper presents a framework for quantifying gender bias through Group Bias ($B_{gr}$) and Individual Bias ($B_{in}$) metrics. Furthermore, we introduce a method to trace these biases to specific training triples using influence functions. By deriving a closed-form approximation of the Hessian matrix for the TransE algorithm, we enable efficient, large-scale bias diagnosis and mitigation. Our results demonstrate that targeted triple removal guided by influence functions significantly outperforms random data reduction in mitigating gender bias.
Introduction
Knowledge Bases (KBs) such as Freebase and Wikidata serve as critical infrastructure for modern AI, encoding millions of facts as relational triples. However, when these structured facts are transformed into low-dimensional vector embeddings, they frequently encapsulate and amplify societal biases. A primary challenge in addressing this issue is the lack of robust measurement tools and the computational difficulty of identifying the specific data points responsible for biased model behavior. This research addresses these gaps by providing a surgical diagnostic toolset for Knowledge Base embeddings.
Methodology
The authors propose two primary metrics to quantify bias:
- Group Bias ($B_{gr}$): Measures the disparity in link prediction error between genders for a specific group (e.g., occupations).
- Individual Bias ($B_{in}$): A counterfactual metric that evaluates the change in prediction error for an entity if its gender attribute were flipped while keeping all other relations constant.
To trace these biases back to training data, the study utilizes influence functions, which estimate the impact of removing a specific training triple on the model's parameters. To overcome the $O(n)$ computational bottleneck of calculating the Hessian matrix, the authors derived a closed-form $O(1)$ approximation specifically for the TransE algorithm, making large-scale analysis feasible.
Data and Benchmarking
The framework was evaluated using the FB5M dataset, a subset of Freebase containing approximately 5 million triples, with a focus on "person" entities and their "profession" relations. To validate the model's internal biases against real-world trends, the authors utilized 2018 U.S. Census Data as a ground-truth benchmark for occupational gender distributions.
Analysis and Key Insights
The research revealed that bias is not merely a function of gender ratios in the training data. For instance, while "animators" exhibited a significant male-to-female ratio, the model's bias leaned slightly toward females, suggesting that the graph's structural connectivity is a critical factor.
Furthermore, the study identified a "popularity buffer": entities with high degrees (many connections) are more stable and less influenced by individual biased triples. Influence was found to stem primarily from direct "1-hop" relations, though "2-hop" indirect influences also contribute to stereotypical associations.
Key Results
The proposed bias measures demonstrated a strong correlation ($r \approx 0.61$) with real-world census data. The influence function's predictions for bias changes were highly accurate, matching actual retraining results with a correlation of $r > 0.95$.
Crucially, the authors demonstrated that targeted de-biasing is highly effective. Removing the top 1% of "pro-bias" triples identified by influence functions reduced bias far more effectively than random triple removal. Additionally, adding balanced "dummy" entities proved to be a viable strategy for bias mitigation.
Limitations and Future Work
Current limitations include the reliance on binary gender definitions due to the underlying KB schemas and the algorithm-specific nature of the $O(1)$ Hessian approximation. Future research should aim to extend these methods to non-binary gender identities, more complex embedding algorithms (e.g., RotatE, ConvE), and other dimensions of bias such as race and nationality.
Summary updated