Improving the Spatial Fairness of Spatiotemporal Sequence Forecasting Service Systems

Spatial unfairness in spatiotemporal forecasting tasks can lead to critical issues, for example, inefficient resource allocation during pandemics or air pollution crises, exacerbating the spread of diseases or increasing exposure to pollution. The existing fair forecasting methods often struggle with spatiotemporal data that lack clear sensitive attributes and exhibit dynamic complexity. A pseudogrouping adversarial machine learning method that combines a spatiotemporal clustering-based pseudogrouping generator with an auxiliary classifier-based adversarial learning framework is proposed in this paper. The proposed pseudogrouping strategy identifies and labels hidden sensitive attributes, enabling fair forecasting models to effectively handle explicit sensitive attributes. Moreover, the existing adversarial learning methods are enhanced in this paper to accommodate the complex dynamics between the features and sensitive attributes contained in the input data. The proposed model is evaluated on three real-world datasets (pandemic, PM2.5, and traffic flow data) and is benchmarked against five existing models using three fairness metrics and three accuracy metrics. The experimental results show that the proposed model significantly outperforms the existing fairness-focused models in terms of accuracy while maintaining fairness, and it markedly improves upon the fairness of accuracy-focused models without sacrificing accuracy. Furthermore, we examine the impact of a fairer forecasting model on the resource allocation process conducted during a pandemic by employing two types of fair resource allocation models. The numerical results derived from the mean–deviation tradeoff model show that using the forecasts obtained from our model can enhance both the efficiency and fairness of allocations when fairness is prioritized during decision-making. When efficiency is prioritized, the forecasting procedure of the proposed model yields improved allocation efficiency, albeit at the cost of reduced fairness. Additionally, results obtained from the α -fairness model indicate that the proposed fair forecasting model consistently outperforms the benchmark across diverse resource allocation objectives, especially when prioritizing one of fairness or efficiency. The findings of this study provide crucial insights into the fairness of forecasting algorithms, illustrating how the introduction of a pseudogrouping strategy and an enhanced adversarial learning framework can significantly improve the fairness of predictive algorithms.