A method of supervised learning from conflicting data with hidden contexts

2021-08-27

AI Summary

This paper addresses the multi-domain input-output relationship conflict arising from latent contextual shifts in open-world environments. We propose LEAF, a method that models training data as drawn from multiple unobserved latent domains, each governed by heterogeneous input-output mappings. LEAF introduces a learnable assignment function that dynamically routes samples to domain-specific predictors; we derive its closed-form solution via a variant of the EM algorithm and establish a theoretical link between this assignment mechanism and generalization error bounds. To our knowledge, LEAF is the first framework to formally characterize and resolve conflicts induced by multiple latent domains, jointly optimizing both assignment and prediction modules. Extensive experiments on synthetic benchmarks and real-world conflict-prone tasks—including multi-source annotation and cross-scenario perception—demonstrate significant improvements over ERM and state-of-the-art domain adaptation methods, validating both theoretical soundness and practical robustness.

Abstract

Conventional supervised learning assumes a stable input-output relationship. However, this assumption fails in open-ended training settings where the input-output relationship depends on hidden contexts. In this work, we formulate a more general supervised learning problem in which training data is drawn from multiple unobservable domains, each potentially exhibiting distinct input-output maps. This inherent conflict in data renders standard empirical risk minimization training ineffective. To address this challenge, we propose a method LEAF that introduces an allocation function, which learns to assign conflicting data to different predictive models. We establish a connection between LEAF and a variant of the Expectation-Maximization algorithm, allowing us to derive an analytical expression for the allocation function. Finally, we provide a theoretical analysis of LEAF and empirically validate its effectiveness on both synthetic and real-world tasks involving conflicting data.

Problem

• Supervised learning with hidden contexts
• Handling data from multiple unobservable domains
• Effective allocation of conflicting data to models

Innovation

• LEAF method allocation function
• Expectation-Maximization algorithm variant
• Handles conflicting hidden contexts data