On the Application of Inductive Machine Learning Tools to Geographical Analysis

Abstract

Inductive machine learning tools, such as neural networks and decision trees, offer alternative methods for classification, clustering, and pattern recognition that can, in theory, extend to the complex or “deep” data sets that pervade geography. By contrast, traditional statistical approaches may fail, due to issues of scalability and flexibility. This paper discusses the role of inductive machine learning as it relates to geographical analysis. The discussion presented is not based on comparative results or on mathematical description, but instead focuses on the often subtle ways in which the various inductive learning approaches differ operationally, describing (1) the manner in which the feature space is partitioned or clustered, (2) the search mechanisms employed to identify good solutions, and (3) the different biases that each technique imposes. The consequences arising from these issues, when considering complex geographic feature spaces, are then described in detail. The overall aim is to provide a foundation upon which reliable inductive analysis methods can be constructed, instead of depending on piecemeal or haphazard experimentation with the various operational criteria that inductive learning tools call for. Often, it would appear that these criteria are not well understood by practitioners in the geographic sphere, which can lead to difficulties in configuration and operation, and ultimately to poor performance.