Multi-scale Traffic Pattern Bank for Cross-city Few-shot Traffic Forecasting

TL;DR

This work tackles cross-city few-shot traffic forecasting under data scarcity by uncovering similarities in multi-scale traffic patterns across cities. It introduces MTPB, a four-module framework that pre-trains a spatial-temporal encoder on data-rich cities, builds a multi-scale traffic pattern bank through clustering, and uses pattern aggregation with a self-expressive graph reconstruction to guide forecasting. Meta-learning via Reptile initializes Pattern Aggregation and Forecasting for rapid adaptation to a target city with limited data, yielding state-of-the-art results on four real-world datasets. The approach offers practical impact for ITS in cities with sparse sensor coverage and provides a blueprint for transferring robust cross-city knowledge to other time-series domains.

Abstract

Traffic forecasting is crucial for intelligent transportation systems (ITS), aiding in efficient resource allocation and effective traffic control. However, its effectiveness often relies heavily on abundant traffic data, while many cities lack sufficient data due to limited device support, posing a significant challenge for traffic forecasting. Recognizing this challenge, we have made a noteworthy observation: traffic patterns exhibit similarities across diverse cities. Building on this key insight, we propose a solution for the cross-city few-shot traffic forecasting problem called Multi-scale Traffic Pattern Bank (MTPB). Primarily, MTPB initiates its learning process by leveraging data-rich source cities, effectively acquiring comprehensive traffic knowledge through a spatial-temporal-aware pre-training process. Subsequently, the framework employs advanced clustering techniques to systematically generate a multi-scale traffic pattern bank derived from the learned knowledge. Next, the traffic data of the data-scarce target city could query the traffic pattern bank, facilitating the aggregation of meta-knowledge. This meta-knowledge, in turn, assumes a pivotal role as a robust guide in subsequent processes involving graph reconstruction and forecasting. Empirical assessments conducted on real-world traffic datasets affirm the superior performance of MTPB, surpassing existing methods across various categories and exhibiting numerous attributes conducive to the advancement of cross-city few-shot forecasting methodologies. The code is available in https://github.com/zhyliu00/MTPB.

Figures (10)

• Figure 1: Multi-scale Traffic Patterns. (a)$\sim$(d) A rapid increase in speed is indicative of the future speed continuing to increase. (e)$\sim$(h) A nearby drop can lead to a subsequent rapid increase in the future. (i)$\sim$(l) If the speed has recently undergone a significant increase, the future speed is likely to exhibit fluctuations.
• Figure 2: Diagrams of MTPB. 1) In the Pre-training stage, a traffic patch encoder is pre-trained by the data of source cities. 2) In the Multi-scale Pattern Generation stage, source city traffic patches are processed by the pre-trained encoder, and the output embeddings are clustered to form the traffic pattern bank. 3) In the Pattern Aggregation stage, the traffic pattern bank is queried by the input traffic patches, and an adjacency matrix is reconstructed. 4) In the Forecasting stage, the metaknowledge along with the short-term and long-term models forecasts the future traffic.
• Figure 3: Ablation Study. We remove queried metaknowledge, ST-Decoder in Pretraining, Multi-scale Pattern, and Graph Reconstruction module separately and evaluate the performance.
• Figure 4: The performance and Silhouette score of traffic patterns w.r.t. different K in Shenzhen.
• Figure 5: The performance of different STmodels trained in meta-learning framework Reptile and trained in our framework MTPB.

Theorems & Definitions (2)

• Definition 1
• Definition 2