FairTTTS: A Tree Test Time Simulation Method for Fairness-Aware Classification

TL;DR

FairTTTS addresses bias in decision-tree classifiers by applying a Monte Carlo, distance-aware traversal at inference to adjust paths involving protected attributes as a post-processing step. Building on TTTS, it introduces a fairness-oriented flip probability controlled by the parameter α and distance to decision thresholds, evaluated on seven benchmarks with metrics including Equalized Odds Difference and Disparate Impact, while achieving occasional accuracy gains. The approach preserves the original model by avoiding retraining, but incurs higher inference time due to Monte Carlo simulations; hyperparameters S=100, p_{max}=0.1, and α=9.0 were used. Overall, FairTTTS demonstrates substantial fairness improvements across diverse datasets and protected attributes with competitive or better accuracy, and is accompanied by publicly available code for reproducibility.

Abstract

Algorithmic decision-making has become deeply ingrained in many domains, yet biases in machine learning models can still produce discriminatory outcomes, often harming unprivileged groups. Achieving fair classification is inherently challenging, requiring a careful balance between predictive performance and ethical considerations. We present FairTTTS, a novel post-processing bias mitigation method inspired by the Tree Test Time Simulation (TTTS) method. Originally developed to enhance accuracy and robustness against adversarial inputs through probabilistic decision-path adjustments, TTTS serves as the foundation for FairTTTS. By building on this accuracy-enhancing technique, FairTTTS mitigates bias and improves predictive performance. FairTTTS uses a distance-based heuristic to adjust decisions at protected attribute nodes, ensuring fairness for unprivileged samples. This fairness-oriented adjustment occurs as a post-processing step, allowing FairTTTS to be applied to pre-trained models, diverse datasets, and various fairness metrics without retraining. Extensive evaluation on seven benchmark datasets shows that FairTTTS outperforms traditional methods in fairness improvement, achieving a 20.96% average increase over the baseline compared to 18.78% for related work, and further enhances accuracy by 0.55%. In contrast, competing methods typically reduce accuracy by 0.42%. These results confirm that FairTTTS effectively promotes more equitable decision-making while simultaneously improving predictive performance.

Figures (4)

• Figure 1: An illustrative example of the FairTTTS approach. At nodes involving protected attributes (shown at the second depth right node), unprivileged samples are directed toward the unfavorable class. By increasing the flipping probability by a factor $\alpha$ at these nodes, FairTTTS probabilistically redirects more of these samples toward favorable leaf nodes, effectively providing greater opportunities to the unprivileged group.
• Figure 2: Accuracy vs. Equalized Odds Across Datasets, Protected Attributes and Methods for Random Forest Architecture. Each plot represents a particular experiment, showing the performance of different methods in different colors. Notably, in seven out of eight experiments, FairTTTS achieves the best fairness, and in all eight experiments, it performs better than the baseline in terms of fairness outcomes.
• Figure 3: Accuracy vs. Equalized Odds Difference across datasets, protected attributes, and methods for Decision Tree architectures. Each plot represents a particular experiment, showing the performance of different methods in distinct colors. Notably, in seven out of eight experiments, FairTTTS achieves the highest accuracy. In six out of eight experiments, fairness improves compared to the baseline, and in four out of eight experiments, fairness improves compared to related work.
• Figure 4: Sensitivity of FairTTTS to the Fairness Adjustment Factor $\alpha$: The figure illustrates the effects of varying $\alpha$ on Accuracy (solid blue lines, left vertical axis) and Equalized Odds Difference (dashed orange lines, right vertical axis). Trend lines (red for Accuracy and green for EOD) highlight the general patterns, showing that moderate $\alpha$ values often achieve improved fairness (lower EOD) without significantly compromising Accuracy, as indicated by the downward slopes of the trend lines.