Abstract
Machine learning models often inherit and amplify societal biases present in their training data. This paper presents a novel multi-objective learning framework that treats bias reduction as a secondary task with a negative objective. By penalizing the model for successfully predicting socially sensitive associations, the framework induces a data representation that is invariant to identity attributes such as race and gender. Evaluation on the Equity Evaluation Corpus (EEC) shows that this approach significantly enhances fairness while maintaining or even improving predictive accuracy compared to standard debiasing techniques.
Introduction
Artificial Intelligence (AI) systems act as mirrors to the data they consume. Because historical data is frequently saturated with human prejudices—such as gender stereotypes in professional roles or racial biases in criminal justice—machine learning models tend to "pick up" and perpetuate these harmful patterns. This phenomenon, termed "cognitive bias," can lead to AI systems disproportionately associating specific demographics with negative traits or risks.
Existing solutions, such as manually removing sensitive features or "debiasing" word embeddings, are often insufficient. They may fail to address implicit biases that can be inferred from non-sensitive features, or they may not account for the complex ways labels interact with identity attributes during the final prediction stage. Furthermore, even when training data is perfectly balanced, models can still produce biased outputs if they rely on pre-trained components (like Google News word vectors) that are already "poisoned" with social stereotypes.
Methodology: Bias-Aware Multi-Objective Learning
The authors propose a Bias-Aware Multi-Objective Learning framework. The core principle is to train the model on two tasks simultaneously:
- The Primary Task: Predicting the correct label (e.g., "What emotion is expressed in this text?").
- The Bias Task: Predicting "biased associations" (e.g., "Is this a female-fear association?").
The Negative Objective Function
In a standard multi-task setup, the goal is to maximize performance across all tasks. In this framework, however, the authors introduce a negative sign in the loss function for the bias tasks. The objective function essentially seeks to maximize the likelihood of correct primary predictions while minimizing the likelihood of identifying biased associations. By penalizing the model for being "good" at identifying stereotypes, the framework forces the underlying neural network to find an abstract representation of the data that is "blind" to those specific harmful correlations. Crucially, identity attributes (race, gender) are used only during the training phase to define these bias tasks; they are not included as input features for the model itself.
Experimental Evaluation
The framework was tested across several datasets to evaluate its robustness:
- Synthetic 2D Data: A Gaussian Mixture Model (GMM) used to visualize how the decision boundary shifts when the model is forced to ignore a specific "biased" region of the feature space.
- Equity Evaluation Corpus (EEC): A real-world dataset containing 8,640 sentences. Each sentence is associated with an emotion (Fear, Anger, Joy, Sadness) and a social identity (Gender and Race) represented through names and pronouns.
- Subsampled Biased Sets (SS-1 & SS-2): Intentionally skewed versions of the EEC used to test if the model could overcome explicit data imbalances (e.g., where women are more frequently associated with fear).
Key Results and Insights
The models were evaluated using Accuracy (A), Fairness (F) (measured by the uniformity of predictions across identity groups), and $\gamma$ (the harmonic mean of the two).
- Superior Fairness: The proposed "Bias-Aware" models significantly outperformed standard models and debiased embedding models (like Bolukbasi’s DWE) in fairness metrics. For example, in the "Fear vs. Gender" task, the proposed model achieved a fairness score of ~0.25 compared to 0.00 for several baselines.
- Accuracy Gains: Surprisingly, the bias-aware models often achieved higher accuracy than the baselines. The authors suggest this is because the model learns to ignore "noisy" biased signals in the word embeddings, focusing instead on the actual semantic content of the sentences.
- Multi-Bias Mitigation: The "Bias-Awr-Joint" model successfully reduced multiple biases (gender and race) simultaneously within a single predictive framework.
- Balanced Data Limitations: A striking finding was that bias-agnostic models still showed significant gender and racial bias even when trained on the perfectly balanced EEC dataset, highlighting the need for algorithmic intervention.
Conclusion and Future Work
This research demonstrates that treating fairness as a trade-off in a multi-objective learning framework can effectively mitigate cognitive bias in AI. By fixing the model during training rather than just pre-processing the data, the system can account for specific posterior distributions of labels.
Limitations include the current need for researchers to manually define which associations are "biased" and the focus on binary identity attributes. Future work will explore the automatic discovery of "bias subspaces" and the inclusion of non-binary and intersectional social identities.
Summary updated