Abstract
This paper introduces a graph-based framework that transforms standard datasets into "Similarity Networks" to tackle the persistent trade-off between model accuracy and fairness. By mapping data instances into a similarity feature space and tuning them with specialized kernels, the authors demonstrate significant improvements in debiasing classification tasks, imputing missing data, and performing fairness-aware data augmentation. Experimental results on a Human Resources dataset show a substantial increase in F1-scores alongside a marked reduction in fairness gaps.
Introduction
Machine learning models are fundamentally limited by the quality of their training data, which often reflects deep-seated social biases. Traditional debiasing techniques—such as removing sensitive features like race or gender or simple class rebalancing—frequently act as blunt instruments that degrade a model's predictive accuracy. Furthermore, these methods struggle with intersectionality, where multiple sensitive attributes overlap.
To address these challenges, this research proposes a more nuanced way to represent data relationships. By adjusting the "resolution" of how instances relate to one another through similarity networks, it is possible to identify a configuration where a model is both high-performing and socially responsible.
Methodology
The core of the research is a three-step methodological pipeline designed to create a robust Similarity Network:
- Similarity Mapping: Depending on the data type, different metrics calculate the proximity between data points. For mixed numerical and categorical tabular data, the authors utilize Gower Distance. For text-heavy data, Natural Language Processing (NLP) techniques like word2vec create embeddings in a latent space.
- Kernel Tuning: To refine the network, two specialized kernels are applied:
- Scaled Exponential Kernel ($E_k$): Captures local topological characteristics.
- Random Walk Kernel ($RW_k$): Captures global structure by simulating paths through the network, strengthening relevant similarities while diminishing noise.
- Network Representation: The final output is a weighted graph $N = (V, E)$, where nodes represent data instances and edges represent the tuned similarity between them.
This secondary "similarity dataset" serves three primary applications: classification (training on similarity features), data imputation (filling missing values via neighbor aggregation), and data augmentation (generating synthetic points to balance underrepresented groups).
Key Insights and Interpretability
A central finding of this work is that biased datasets often exhibit low "classification complexity," allowing models to find biased "shortcuts." By mapping data to a similarity network, the authors intentionally increase complexity, forcing the model to identify deeper, more equitable patterns.
Unlike many "black-box" debiasing algorithms, the Similarity Network is inherently interpretable. Researchers can visualize instance links and utilize tools like SHAP (SHapley Additive exPlanations) to determine exactly which features and neighbors are driving specific predictions.
Experimental Results
The methodology was evaluated using a public Human Resources (HR) dataset containing CV details for 301 employees, with the task of predicting "Pay Rate" categories—a task prone to gender and ethnic bias. The results demonstrated marked improvements:
- Accuracy: The F1-weighted score for a Random Forest classifier rose from 0.666 on original data to 0.944 using similarity-mapped data.
- Fairness (Equal Opportunity): The gap in True Positive Rates between privileged and unprivileged groups was reduced from 0.3 to 0.2.
- Fairness (Equal Mis-opportunity): The False Positive Rate gap was slashed from 0.17 to just 0.02.
- Augmentation: Fairness-aware augmentation improved class balance by 60% and halved the Equal Opportunity gap.
Conclusion
While the results on the HR dataset are promising, the methodology's performance across different domains remains to be validated. Future work will focus on testing the framework against a wider variety of datasets, automating the kernel selection process for the accuracy-fairness trade-off, and exploring applications for real-time streaming data.
Summary updated