Abstract
This short paper improves GeoAggregator, a Transformer-based deep learning model for geospatial tabular data, by optimizing the data loading and forward-pass pipeline to enhance computational efficiency and by incorporating model ensembling and a post-hoc explanation function based on the GeoShapley framework. Experiments on synthetic datasets show improved prediction accuracy and inference speed compared with the original implementation, while the explanation results demonstrate that the model can effectively capture inherent spatial effects.
Type
Publication
The 8th ACM SIGSPATIAL International Workshop on AI for Geographic Knowledge Discovery (GeoAI ‘25). ACM, New York, NY, USA. https://doi.org/10.1145/3764912.3770843
Rui Deng
PhD Researcher
