Decomposing spatio-temporal heterogeneity: Matrix-informed ensemble learning for interpretable prediction

•A matrix-informed ensemble learning for interpretable prediction was proposed.•The spatio-temporal heterogeneity was expressed by matrix decomposition.•The spatio-temporal score vectors were designed to explain the ensemble process.•The interpretability and prediction accuracy of proposed method were validated. Spatio-temporal prediction aims to forecast location or time-related changes in the physical world with disciplinary knowledge and historical data. Ensemble learning strategies that integrate multiple base learners can leverage the advantages of different models and thus gain wide attention in the field of spatio-temporal prediction. However, existing methods often ignore the unified expression of spatial and temporal heterogeneity in the ensemble process and fail to clarify the mechanisms of model integration under these constraints, limiting the predictive ability and the interpretability of ensemble models. Therefore, this study proposed a Matrix-Informed Ensemble Learning Method (MI-EL) for interpretable spatio-temporal prediction. The core idea of this method is to decompose the spatio-temporal heterogeneous ensemble weight matrix into the multiplication of the spatial and temporal factor matrix. By constructing a spatio-temporal embedding learning module, it utilizes spatial associations and temporal attributes of the samples to solve the spatial and temporal factor matrices, thereby achieving a unified expression of spatial and temporal heterogeneity in the ensemble process. On this basis, interpretable spatial and temporal score vectors are constructed for explicitly expressing the influence intensities and response rules of different base learners under conditions of spatial and temporal heterogeneity. Experiments on traffic flow, traffic speed and air quality prediction tasks show that the proposed method outperforms eight existing ensemble methods in the prediction accuracies at different time steps. Additionally, it effectively identifies the performance patterns of base learners at different spatio-temporal units by assigning higher spatio-temporal scores to better-performing base learners, thereby achieving superior prediction results.