关键词: 可预测性, 空间面板数据, 空间依赖性, 空间异质性, 地理相似性, 熵

Abstract:

Spatial panel data, characterized by the regularity of information in both spatial and temporal dimensions, are commonly used to record the spatial-temporal evolution of geographic phenomena, and are the mainstream data structure for spatial-temporal prediction research. Especially in the era of artificial intelligence centered on neural networks, spatial panel data can be input into intelligent models such as convolutional networks and recurrent networks without additional processing, which has the advantages of non-destructive information and convenient computation, and is commonly used in prediction research in the fields of human activities and transportation. However, the existing research focuses on the enhancement of modeling methods, and the theoretical issues of whether the data itself can be predicted, to what extent it can be predicted, and how it should be predicted are seldom addressed. There is entropy-based predictability theory in the field of information and statistics, which is widely used in time series analysis, but it neglects geo-spatial effects such as spatial dependence, spatial heterogeneity and geographic similarity in the spatial panel data and its influence mechanism on the prediction potential and modeling approach, which leads to inaccurate assessment results. In this regard, based on the existing predictability assessment theory, this paper proposes the geographic entropy theory and method, including neighborhood transfer entropy, cross-space entropy and cross-region entropy, taking into account the influence mechanism of geo-spatial effects, so as to quantitatively assess the predictability of spatial panel data from the different aspects of feature learning, parameter training, and application testing and to provide theoretical basis for the study of spatial-temporal prediction, spatial neighborhood learning, local model construction, and transfer and generalization strategy.

Key words: predictability, spatial panel data, spatial dependence, spatial heterogeneity, geographic similarity, entropy