Spatial regression-based transfer learning for prediction problems

TL;DR

This paper introduces a novel transfer learning method that pre-trains spatial-dependent processes to improve prediction accuracy in spatial tasks with limited local data, demonstrating enhanced stability and accuracy.

Contribution

It proposes a new spatial regression-based transfer learning approach that explicitly incorporates local spatial dependence, improving prediction performance in small-sample scenarios.

Findings

Improved land price prediction accuracy.

Enhanced stability in crime prediction.

Outperformed conventional methods in experiments.

Abstract

Although spatial prediction is widely used for urban and environmental monitoring, its accuracy is often unsatisfactory if only a small number of samples are available in the study area. The objective of this study was to improve the prediction accuracy in such a case through transfer learning using larger samples obtained outside the study area. Our proposal is to pre-train latent spatial-dependent processes, which are difficult to transfer, and apply them as additional features in the subsequent transfer learning. The proposed method is designed to involve local spatial dependence and can be implemented easily. This spatial-regression-based transfer learning is expected to achieve a higher and more stable prediction accuracy than conventional learning, which does not explicitly consider local spatial dependence. The performance of the proposed method was examined using land price and crime predictions. These results suggest that the proposed method successfully improved the accuracy and stability of these spatial predictions.