Group-wise oracle-efficient algorithms for online multi-group learning

2024-06-07

Neural Information Processing Systems

AI Summary

This paper studies online multi-group learning, where groups correspond to implicit subsets of the context space—such as overlapping subpopulations defined by complex demographic attributes—and are too numerous to enumerate explicitly. The goal is to achieve sublinear group-level regret. We propose the first oracle-efficient online algorithm that relies solely on an optimization oracle—without explicit group enumeration—and applies to i.i.d., smoothed adversarial, and adversarial transductive settings. Leveraging a group-aware regret decomposition, smoothed analysis, and a transductive learning framework, we establish sublinear $O(sqrt{T})$ regret bounds simultaneously for the overall policy and every individual group—significantly improving upon prior enumeration-based approaches. The algorithm’s time complexity is polynomial in the number of oracle calls, ensuring practical scalability.

Technology Category

• Machine Learning: Online Learning & Bandits
• Search and Optimization: Learning to Search
• Intelligent Robots: Learning & Optimization for ROB

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs
• Economics, Online Markets and Human Computation: Fairness, privacy, and diversity in economic environments

Abstract

We study the problem of online multi-group learning, a learning model in which an online learner must simultaneously achieve small prediction regret on a large collection of (possibly overlapping) subsequences corresponding to a family of groups. Groups are subsets of the context space, and in fairness applications, they may correspond to subpopulations defined by expressive functions of demographic attributes. In contrast to previous work on this learning model, we consider scenarios in which the family of groups is too large to explicitly enumerate, and hence we seek algorithms that only access groups via an optimization oracle. In this paper, we design such oracle-efficient algorithms with sublinear regret under a variety of settings, including: (i) the i.i.d. setting, (ii) the adversarial setting with smoothed context distributions, and (iii) the adversarial transductive setting.

Problem

• Online multi-group learning with large group families
• Achieving sublinear regret via oracle-efficient algorithms
• Handling overlapping groups without explicit enumeration

Innovation

• Oracle-efficient algorithms for large group families
• Sublinear regret in i.i.d. and adversarial settings
• Optimization oracle access without explicit enumeration