Efficient Machine Learning for Semantic Inference on Spatial Networks

Geo-AI is a discipline that leverages both artificial intelligence and geographical information systems. A practical application of Geo-AI for problem solving is the task of inferring semantics for spatial networks, otherwise known as semantic inference, which involves inferring the semantic type of spatial entities such as labelling the type of a road or the use of a building. An application of the semantic inference task is to address the open research problem of improving data quality in crowd-sourced spatial databases. Herein, the data quality problem can be tackled through automatically predicting their semantics using machine learning techniques. However the performance of the machine learning models are impacted by some issues. 1) Training a machine learning model with acceptable results requires a lot of training data which may not be available or affordable. 2) Further, the models are prone to overfit on the training data, which limits their capacity to generalise sufficiently on unseen data. 3) In the same vein, the machine learning algorithms make assumptions about the data which could impact model performance. For example, some machine learning methods assume independence or homophily in the data which is not necessarily guaranteed for geo-spatial data. These assumptions - when contradicted - could be detrimental to model performance. In light of the aforementioned issues, it becomes clear that the efficiency of the machine learning algorithms for semantic inference needs to be investigated. In this thesis, we explore the development of efficient machine learning techniques for semantic inference on spatial networks. We make two arguments: (a) that leveraging existing data from one domain to train a machine learning model for a task in another domain could improve model performance, where a domain could be a city or any place within an officially recognised boundary (b) that the data representations used for a machine learning task impact the efficiency of the model, thus evaluating the robustness of representations is critical. Consequently, we make our case through evidence-based experiments. We formulate research questions that are tested using data collected from OpenStreetMap. In this thesis, we make the following contributions: 1) A method for the efficient development of graph machine learning models for semantic inference of spatial networks that out-performs state-of-the art methods. 2) A neural model for training transferable graph neural networks for spatial networks that mitigates negative transfer and improves transfer gain. 3) An end-to-end model for learning on heterogeneous representations of spatial networks. The contributions made by this thesis will benefit the advancement of Geo-AI by offering insights into developing efficient inference models.