cs.LGMar 19, 2026
SaveA Model Ensemble-Based Post-Processing Framework for Fairness-Aware Prediction
Organizations: School of Computer Science and Statistics, Trinity College Dublin, Dublin, Ireland.
Abstract
Striking an optimal balance between predictive performance and fairness continues to be a fundamental challenge in machine learning. In this work, we propose a post-processing framework that facilitates fairness-aware prediction by leveraging model ensembling. Designed to operate independently of any specific model internals, our approach is widely applicable across various learning tasks, model architectures, and fairness definitions. Through extensive experiments spanning classification, regression, and survival analysis, we demonstrate that the framework effectively enhances fairness while maintaining, or only minimally affecting, predictive accuracy.
Figures & tables
| Task | Dataset | Size | # Features | Sensitive Attribute(s) | Non-sensitive Feature (Fair Model) |
| Binary Classification | COMPAS | 7214 | 6 | gender, race | c_charge_degree |
| Adult | 46447 | 21 | gender, race | education-num | |
| Heart | 319795 | 17 | gender | BMI | |
| German Credit | 1000 | 33 | gender | Duration | |
| Regression | Insurance | 1338 | 6 | gender | BMI |
| Survival Analysis | WHAS | 481 | 9 | gender | lenstay |
Table 1: Summary of Datasets, Sensitive Attributes, and Non-sensitive Features Used in the Fairness Models
Figure 1: Performance and fairness trade-offs on the Adult dataset across varying performance models and sensitive-attribute settings. The figure consists of 20 panels arranged in a 5 4 grid. Rows represent different model configurations: (Row 1) 1-pretrained Mixture , (Row 2) 1-pretrained MoE , (Row 3) 2-pretrained Mixture , (Row 4) 2-pretrained MoE , and (Row 5) the FRAPPÉ baseline. Columns correspond to experimental settings: (Col 1) RF as performance model, sensitive attribute: sex ; (Col 2) MLP as performance model, sensitive attribute: sex ; (Col 3) RF as performance model, sensitive attributes: sex + race ; (Col 4) MLP as performance model, sensitive attributes: sex + race . Tested values: (Col 1) [0.01, 0.5, 1, 5, 10, 100, 200, 300, 500], (Col 2) [0.01, 0.05, 1, 5, 10, 100, 500], (Col 3) [0.01, 0.05, 1, 5, 10, 100], (Col 4) [0.01, 0.05, 1, 5, 10, 100, 500].
Figure 2: Performance and fairness trade-offs on the COMPAS dataset across varying performance models and sensitive-attribute settings. The figure consists of 20 panels arranged in a 5 4 grid. Rows correspond to model configurations: (Row 1) 1-pretrained Mixture , (Row 2) 1-pretrained MoE , (Row 3) 2-pretrained Mixture , (Row 4) 2-pretrained MoE , and (Row 5) the FRAPPÉ baseline. Columns correspond to experimental settings: (Col 1) RF as performance model, sensitive attribute: gender ; (Col 2) MLP as performance model, sensitive attribute: gender ; (Col 3) RF as performance model, sensitive attributes: gender + race ; (Col 4) MLP as performance model, sensitive attributes: gender + race . Tested values: (Cols 1–2) [1, 5, 10, 100, 200, 300, 500, 700, 1000]; (Cols 3–4) [0.01, 0.05, 0.1, 0.2, 0.5, 1, 5, 10, 30, 50].
Figure 3: Performance and fairness trade-offs on the Heart dataset across different model configurations and performance models. The figure contains 10 panels arranged in a 5 × 2 grid. Rows correspond to model configurations: (Row 1) 1-pretrained Mixture , (Row 2) 1-pretrained MoE , (Row 3) 2-pretrained Mixture , (Row 4) 2-pretrained MoE , and (Row 5) the FRAPPÉ baseline. Columns correspond to performance models: (Col 1) RF with sensitive attribute gender ; (Col 2) MLP with sensitive attribute gender . Tested values: (RF) [0.01, 0.5, 1, 5, 10, 20, 50, 100, 200, 300, 500, 600]; (MLP) [0.01, 0.5, 1, 5, 10].
Figure 4: Performance and fairness trade-offs on the German dataset across different model configurations and performance models. The figure contains 10 panels arranged in a 5 × 2 grid. Rows correspond to model configurations: (Row 1) 1-pretrained Mixture , (Row 2) 1-pretrained MoE , (Row 3) 2-pretrained Mixture , (Row 4) 2-pretrained MoE , and (Row 5) the FRAPPÉ baseline. Columns correspond to performance models: (Col 1) RF with sensitive attribute gender ; (Col 2) MLP with sensitive attribute gender . Tested values: [0.01, 0.1, 0.5, 1, 5, 10, 20].
| Dataset | Base | Method | Perf ( ) | DP ( ) | EO ( ) | Variant | |
| Adult | RF | ROC | 0.7752 | 0.0468 | 0.1329 | New A | New A |
| CalibratedEqOddsPostprocessing | 0.8327 | 0.2575 | 0.6033 | New A | New A | ||
| Reductions | 0.8404 | 0.0119 | 0.3163 | New A | New A | ||
| HGR (DP-targeted) | New A | New A | New A | New A | New A | ||
| HGR (EO-targeted) | New A | New A | New A | New A | New A | ||
| Ours — Best-Perf | 0.8475 | 0.1769 | 0.0885 | 1-pretrained, Mixture | 0.01 |
Table 2: Binary — four datasets, two base models (RF/MLP). Perf: Accuracy ( ). Fairness reports both DP and EO ( ). Baselines are single-point; ours reports two points (Best-Perf / Best-Fair) selected across our four internal variants.
Figure 5: Performance and fairness trade-offs on the Insurance dataset. The figure contains five panels. Shown from left to right: 1-pretrained Mixture , 1-pretrained MoE , 2-pretrained Mixture , 2-pretrained MoE , and the FRAPPÉ baseline. The performance model used is Random Forest with sensitive attribute gender . Tested values: [0.001, 0.005, 0.01, 0.02, 0.03, 0.04, 0.05, 0.1, 0.5, 1, 5, 10, 20, 50].
Figure 6: Performance and fairness trade-offs on the WHAS survival dataset. The figure contains four panels arranged in two rows. Shown from left to right, top to bottom: 1-pretrained Mixture , 1-pretrained MoE , 2-pretrained Mixture , and 2-pretrained MoE . The performance model used is Random Survival Forest with sensitive attribute gender . Baseline (Zhao and Ng, 2023 Zhao and Ng (2023) ) : C-index =0.7703, IBS =0.1783, AUC =0.8644, fairness =0.036. Tested values: (a) [0.01, 0.05, 0.1, 1, 5, 10, 15, 20, 100, 500]; (b) [1, 5, 10, 50, 100, 200, 300, 500]; (c) [0.01, 0.05, 0.1, 1, 5, 10, 15, 20, 100, 200, 500]; (d) [0.01, 0.05, 0.1, 1, 5, 10].
Figure 7: Performance and fairness trade-offs on the Employee survival dataset. The figure contains four panels arranged in two rows (2 × 2 layout). Shown from left to right, top to bottom: 1-pretrained Mixture , 1-pretrained MoE , 2-pretrained Mixture , and 2-pretrained MoE . The performance model used is Random Survival Forest with sensitive attribute gender . Baseline (Zhao and Ng, 2023 Zhao and Ng (2023) ) : C-index =0.6602, IBS =0.1783, AUC =0.7423, fairness =0.0088. Tested values: (a) [0.001, 0.1, 1, 5, 20, 50]; (b) [0.001, 0.005, 0.1, 0.5, 1, 5, 20, 50, 100]; (c) [0.01, 0.05, 0.1, 0.5, 1, 5]; (d) [0.01, 0.05, 0.1, 1, 5, 10].
Figure 8: Effectiveness of our method under increased model capacity using 3-layer MLPs. Each pair compares the performance of the proposed MoE model with a single pre-trained backbone against the FRAPPÉ baseline on the Adult and COMPAS datasets. Tested values: Adult — [1, 5, 10, 100, 200, 300, 500]; COMPAS — [1, 5, 10, 100, 200, 300, 500, 700, 1000].
Figure 9: t-SNE visualizations of instance-level weights on the Adult and COMPAS datasets under both single- and multi-sensitive attribute settings. In each subfigure, the first two rows ((a)–(h)) correspond to Adult , and the last two ((i)–(p)) to COMPAS . For each dataset, the first row uses Random Forest (RF) as the performance model, and the second uses a 1-layer Multi-Layer Perceptron (MLP). Each pair of columns shows overall (left) and subgroup-based (right) visualizations of instance-level weights. The left half of each row corresponds to the 1-pretrained MoE model, and the right half to the 2-pretrained MoE model.
Figure 10: t-SNE visualizations of instance-level weights from MoE and post-processed MoE models across five datasets under single-sensitive (gender) settings. Random Forest and Random Survival Forest are used as base predictors.