Core Concepts
Performance heterogeneity in spatiotemporal learning can lead to unfair predictions, but FairSTG offers a model-independent solution to enhance fairness and accuracy.
Abstract
The article introduces FairSTG, a framework designed to address performance heterogeneity in spatiotemporal learning for urban applications. It highlights the importance of fair predictions and the risks associated with unfair models. The framework consists of a spatiotemporal feature extractor, fairness recognizer, collaborative feature enhancement, and output module. By identifying challenging samples and compensating for them with well-learned samples, FairSTG aims to improve fairness while maintaining accuracy. Experimental results on four datasets demonstrate the effectiveness of FairSTG in improving prediction fairness.
Stats
DCRNN achieves an MAE variance of 51.61% on METR-LA dataset.
STGCN has an MAPE variance of 15.75% on PEMS-BAY dataset.
MTGNN shows an MAE variance of 47.69% on METR-LA dataset.
D2STGNN demonstrates an MAPE variance of 9.43% on PEMS-BAY dataset.