Research map / Bandits and online learning / Online convex optimization and games
Causal Logistic Bandits with Counterfactual Fairness Constraints
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 CCLB, a primal-dual algorithm for causal logistic bandits with an a priori unknown counterfactual fairness constraint, achieving sublinear regret and violations, with a tunable slack that gives zero cumulative violations.
Problem
Sequential decision-making where decisions must satisfy a counterfactual fairness constraint defined with respect to a specified attribute, under a logistic reward model with unknown parameters and no known safe action set.
Approach
Combines a logistic-bandit optimistic confidence set (loss-based, convex) with primal-dual optimization on a soft long-term counterfactual fairness constraint, so the learner does not need a known safe policy; a tightness parameter epsilon in the dual update trades regret for zero cumulative violations.
Key results
- CCLB attains regret leading term O~(rho n S sqrt(T)) and violations O~(n S sqrt(T) + n^2 S^2 kappa_Z + sqrt(T)), matching the unconstrained OFULog+ leading term up to rho.
- With a tightness parameter epsilon, the violation bound decreases by epsilon*T, so zero cumulative violations are achievable for suitable epsilon (Proposition 4) at a small regret cost.
- In synthetic experiments, CCLB has much lower penalized cumulative regret than GLM-UCB and OFULog+, while unconstrained OFULog+ achieves higher unpenalized cumulative regret.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add counterfactual fairness constraints using primal-dual optimization | Logistic bandits · Constrained causal logistic bandits | reward regret: eO(ρnS√T + ρn²S²κZ + ρ√T) | Helped |
| Add counterfactual fairness constraints using primal-dual optimization | Logistic bandits · Constrained causal logistic bandits | cumulative constraint violations: eO(nS√T + n²S²κZ + √T) | Helped |
| Add a tightness parameter in the dual update (ablation) | CCLB · Constrained causal logistic bandits | cumulative constraint violations: eO(nS√T + n²S²κZ + (1 + ϵ)²√T − ϵT) -> 0 | Helped |
| Add a tightness parameter in the dual update (ablation) | CCLB · Constrained causal logistic bandits | reward regret: ρ√T(1 + ϵ)² | Hurt |
| Add counterfactual fairness awareness to the unconstrained bandit method | Synthetic data · Synthetic data | penalized cumulative regret | Helped |
| Use the tightness-parameter variant ϵ-CCLB (ablation) | CCLB · Synthetic data | penalized cumulative regret | Helped |
| Use the tightness-parameter variant ϵ-CCLB (ablation) | CCLB · Synthetic data | cumulative regret | Hurt |
| Use the tightness-parameter variant ϵ-CCLB (ablation) | CCLB · Synthetic data | cumulative constraint violations | Mixed |
| Account for fairness constraints in action selection | OFULog+ and GLM-UCB · Synthetic data | penalized cumulative regret | Helped |
Surprising findings
- Unconstrained OFULog+ has higher unpenalized cumulative regret than the constrained CCLB, since it ignores fairness and gets rewards from unfair actions that count in unpenalized regret.
Limitations
- Assumes observed confounders (W observed); unobserved confounding not handled
- Constraint violations are only bounded cumulatively, so individual decisions may be unfair
- Regret bound has an extra factor rho from the truncated dual variable, and the zero-violation regime costs regret
Experimental setup
Datasets: synthetic structural causal model data
Benchmarks: synthetic causal logistic bandit environment
Metrics: cumulative regret cumulative constraint violations penalized cumulative regret
Baselines: GLM-UCB OFULog+ CCLB
Builds on: OFULog+ Huang et al. 2022b causal fair bandits Amani et al. 2020 safe GLM-UCB Zhou & Ji 2022 primal-dual kernelized bandits Plecko & Bareinboim extended standard fairness model
Closest papers by meaning
- Causal Bandits: The Pareto Optimal Frontier of Adaptivity, a Reduction to Linear Bandits, and Limitations around Unknown Marginals (ICML 2024 · similarity 0.87)
- Logistic Bandits with $\tilde{O}(\sqrt{dT})$ Regret without Context Diversity Assumptions (NeurIPS 2026 · similarity 0.87)
- Robustly Improving Bandit Algorithms with Confounded and Selection Biased Offline Data: A Causal Approach (AAAI 2024 · similarity 0.87)
- Honor Among Bandits: No-Regret Learning for Online Fair Division (NeurIPS 2024 · similarity 0.86)
- Partial Structure Discovery is Sufficient for No-regret Learning in Causal Bandits (NeurIPS 2024 · similarity 0.86)
- Improved Regret Bounds for Online Fair Division with Bandit Learning (AAAI 2025 · similarity 0.86)
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.