Referee-Meta-Learning for Fast Adaptation of Locational Fairness

TL;DR

This work addresses locational biases in geo-located predictive models and proposes a locational meta-referee (Meta-Ref) that modulates per-location learning rates during model-agnostic meta-learning (MAML) to achieve prediction quality parity across locations. Meta-Ref receives per-location performance, a global fairness baseline, and sample encodings to produce location-specific fairness factors that guide inner-loop updates. A three-phase training framework builds a distribution of spatial tasks and optimizes both predictive accuracy and locational fairness, with dual meta-updates balancing performance and fairness. Experiments on satellite-based crop classification and transportation safety demonstrate improved locational fairness (lower LF and ALF) while preserving overall predictive performance, and analyses highlight the necessity of all three gradient updates for robust transfer to unseen regions.

Abstract

When dealing with data from distinct locations, machine learning algorithms tend to demonstrate an implicit preference of some locations over the others, which constitutes biases that sabotage the spatial fairness of the algorithm. This unfairness can easily introduce biases in subsequent decision-making given broad adoptions of learning-based solutions in practice. However, locational biases in AI are largely understudied. To mitigate biases over locations, we propose a locational meta-referee (Meta-Ref) to oversee the few-shot meta-training and meta-testing of a deep neural network. Meta-Ref dynamically adjusts the learning rates for training samples of given locations to advocate a fair performance across locations, through an explicit consideration of locational biases and the characteristics of input data. We present a three-phase training framework to learn both a meta-learning-based predictor and an integrated Meta-Ref that governs the fairness of the model. Once trained with a distribution of spatial tasks, Meta-Ref is applied to samples from new spatial tasks (i.e., regions outside the training area) to promote fairness during the fine-tune step. We carried out experiments with two case studies on crop monitoring and transportation safety, which show Meta-Ref can improve locational fairness while keeping the overall prediction quality at a similar level.

Figures (6)

• Figure 1: Illustrative examples of spatial tasks for training and testing, their encompassing locations and data points, as well as the data requirements for finetuning/adaption for spatial tasks from test locations.
• Figure 2: An illustration of the training framework to enforce locational fairness with Meta-Ref.
• Figure 3: A pairwise comparison matrix for all methods (This is an example from task set 1 for crop classification with 30 tasks in total).
• Figure 4: Comparison between MAML and Meta-Ref on fairness metrics among different spatial tasks (an example from task set 1 for crop classification with 30 tasks in total).
• Figure A1: Pairwise comparison matrices for all methods for crop classification tasks.

Theorems & Definitions (3)

• Definition 1
• Definition 2
• Definition 3