Artificial Intelligence Review - Special issue on lazy learning
Constructive Incremental Learning from Only Local Information
Neural Computation
Advanced space-time predictive analysis with STAR-BME
Proceedings of the 20th International Conference on Advances in Geographic Information Systems
| Hi-index | 0.00 |
Local regression models are one of the backbones of spatial analytics. Computational scalability of such models can be resolved with a distributed implementation which requires truly local modelling as communication between components is limited or absent at some stages. The use of such models in a streaming context provide further restrictions. A calibration procedure has to be truly incremental, with constant memory and processing time for any sample in a stream. This paper explores a spatially distributed incremental local regression model satisfying these requirements and providing similar functionality in terms of interpretability and modelling accuracy as the widely-used geographically weighted regression. Our experiments were run on a conventional mid-range 8-core server. In the largest scale experiment we processed a stream of 157 million spatially referenced samples simulating power consumption readings taken every 2 hours in a period of 5 months from about 87000 households in a European country. The software implementation we developed for the evaluation and performance analysis is made available as an open source project.