Research Article
Transfer Knowledge Between Cities by Incremental Few-Shot Learning
@INPROCEEDINGS{10.1007/978-3-030-92638-0_15, author={Jiahao Wang and Wenxiong Li and Xiuxiu Qi and Yuheng Ren}, title={Transfer Knowledge Between Cities by Incremental Few-Shot Learning}, proceedings={Collaborative Computing: Networking, Applications and Worksharing. 17th EAI International Conference, CollaborateCom 2021, Virtual Event, October 16-18, 2021, Proceedings, Part II}, proceedings_a={COLLABORATECOM PART 2}, year={2022}, month={1}, keywords={Spatial-temporal prediction Incremental few-shot learning Meta-learning Traffic prediction}, doi={10.1007/978-3-030-92638-0_15} }- Jiahao Wang
Wenxiong Li
Xiuxiu Qi
Yuheng Ren
Year: 2022
Transfer Knowledge Between Cities by Incremental Few-Shot Learning
COLLABORATECOM PART 2
Springer
DOI: 10.1007/978-3-030-92638-0_15
Abstract
The objective of cross-city transfer learning methods focuses on how to effectively transfer knowledge from data-rich cities to help data-scarce cities, and solve the problem that city development levels are quite unbalanced. However, transfer-learning and meta-learning-based spatial-temporal approaches can quickly learn and adapt to (novel-) source cities, but the prior experience in base-source cities will be largely forgotten, i.e., the models may lead to catastrophic forgetting problem on base attributes. In this paper, we proposed an incremental few-shot learning based spatial-temporal model (IFS-STP), which utilized an incremental few-shot learner strives to build a generalized model that can not only transfer learned knowledge from source cities to improve the performance of spatial-temporal prediction in a target city with limited data but also prevent the catastrophic forgetting problem of source cities. We evaluate IFS-STP on traffic prediction tasks and the experience results show that our approach significantly outperforms competitive baseline models.
