Explainable Artificial Intelligence for Crowd Forecasting Using Global Ensemble Echo State Networks

Abstract

Crowd monitoring is a primary function in diverse industrial domains, such as smart cities, public transport, and public safety. Recent advancements in low-energy devices and rapid connectivity have enabled the generation of real-time data streams suitable for crowd-monitoring applications. Crowd forecasting is typically achieved using deep learning models that learn the evolving nature of data streams. The computational complexity, execution time, and opaqueness are inherent challenges of deep learning models that also overlook the latent relationships between multiple real-time data streams for improved accuracy. To address these challenges, we propose the global ensemble echo state network approach for explainable crowd forecasting using multiple WiFi data streams. This approach replaces the random input mapping layer with a clustering layer, allowing the network to learn input projections on cluster centroids. It incorporates an ensemble readout comprising a stack of reservoir layers that provide model explainability. It also learns multiple related time series in parallel to construct a global model that leverage latent relationships across the data streams. This approach was empirically evaluated in a multicampus, mixed-use tertiary education setting. The results of which confirm the effectiveness and interpretability of the proposed approach for industrial applications of crowd forecasting.