MemDA: Forecasting Urban Time Series with Memory-based Drift Adaptation

By Zekun Cai et al

Published on Oct. 21, 2023

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. Related Work
3. Preliminaries
4. Methodology
4.1 Rationale
4.2 Dual-Memory for Drift Embedding
4.3 Meta-Dynamic Network for Drift Adaptation
5. Experiments
6. Conclusion
7. References

Summary

MemDA is a novel approach for forecasting urban time series data with memory-based drift adaptation. It addresses the concept drift problem in urban data forecasting by encoding periodicity and making on-the-fly adjustments to the model based on drift using a meta-dynamic network. The proposed model outperforms existing methods and is adaptable to various prediction backbones. The experiments demonstrate the effectiveness and efficiency of the design.