Research map / Learning theory and algorithms / Fairness
Enhancing Group Fairness in Online Settings Using Oblique Decision Forests
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Aranyani, an ensemble of soft-routed oblique decision trees that enables online group fairness (demographic parity) training by maintaining aggregate node-level statistics instead of storing past instances.
Problem
Enforcing group fairness (demographic parity) during online learning, where instances arrive one at a time, normally requires storing past instances and extra forward/backward passes to estimate fairness gradients.
Approach
Use an ensemble of soft-routed oblique decision trees whose node-level outputs are constrained for demographic parity; the tree's hierarchical structure lets fairness gradients be estimated from running aggregate statistics of node outputs, with no stored samples and no re-forward passes.
Key results
- Aranyani achieves the best accuracy-fairness trade-off against Hoeffding-tree baselines, MLP and Leaf variants across tabular, CivilComments and CelebA setups.
- Compared with the Reservoir variant that stores all samples, Aranyani gives similar accuracy-fairness trade-off on Adult with ~3x computation-time improvement and ~23% memory reduction.
- Fairness and task gradient norms converge over online training on CivilComments, consistent with the theoretical bound.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use aggregate node statistics instead of storing all samples to compute fairness loss (ablation) | Aranyani vs Reservoir · Adult | accuracy-fairness tradeoff | No clear effect |
| Use aggregate node statistics instead of storing all samples to compute fairness loss (ablation) | Aranyani vs Reservoir · Adult | computation time: ∼3x improvement | Helped |
| Use aggregate node statistics instead of storing all samples to compute fairness loss (ablation) | Aranyani vs Reservoir · Adult | memory utilization: ∼23% reduction | Helped |
| Replace the oblique forest with an MLP while retaining the online gradient updates (ablation) | Aranyani · Tabular datasets | demographic parity | Hurt |
| Apply fairness gradients to leaf-level predictions instead of node predictions (ablation) | Aranyani · Tabular datasets | demographic parity | Hurt |
| Replace the oblique forest with a conventional Hoeffding tree (ablation) | Aranyani · CelebA and CivilComments | task accuracy | Hurt |
| Increase tree height (ablation) | Aranyani · Adult | accuracy | Helped |
| Increase tree height (ablation) | Aranyani · Adult | demographic parity | Hurt |
| Increase tree height to 8 (ablation) | Aranyani · Adult | accuracy | Hurt |
| Use Aranyani instead of baseline approaches | Online fair learning methods · Adult, Census, COMPAS, CivilComments, and CelebA | accuracy-fairness tradeoff | Helped |
Surprising findings
- The MLP and Leaf variants (same gradient accumulation without node-level constraints) fail to improve fairness beyond a certain point due to vanishing fairness gradients in layers far from the output.
Limitations
- Binary oblique trees may have limited expressiveness compared with fully connected networks.
- Theoretical gradient convergence bound grows exponentially with tree height (2^{2h}).
- High variance of results on COMPAS, attributed to small dataset size.
Experimental setup
Datasets: UCI Adult Census ProPublica COMPAS CivilComments CelebA
Benchmarks: UCI Adult Census ProPublica COMPAS CivilComments CelebA
Metrics: accuracy demographic parity
Baselines: Hoeffding Tree Adaptive Hoeffding Tree FAHT MLP Aranyani Leaf Aranyani Majority post-processing FERMI Reservoir variant
Builds on: Hoeffding trees FERMI sigmoid tree decomposition Huber loss
Closest papers by meaning
- Re-weighting Based Group Fairness Regularization via Classwise Robust Optimization (ICLR 2023 · similarity 0.90)
- FFB: A Fair Fairness Benchmark for In-Processing Group Fairness Methods (ICLR 2024 · similarity 0.89)
- FACT: A Diagnostic for Group Fairness Trade-offs (ICML 2020 · similarity 0.89)
- FRAPPÉ: A Group Fairness Framework for Post-Processing Everything (ICML 2024 · similarity 0.88)
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.88)
- Some Optimizers are More Equal: Understanding the Role of Optimizers in Group Fairness (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.