Distribution-aware Online Continual Learning for Urban Spatio-Temporal Forecasting

TL;DR

DOST, a novel online continual learning framework tailored for ST data characteristics, is introduced, providing online forecasts within an average of 0.1 seconds and achieving a 12.89% reduction in forecast errors compared to baseline models.

Abstract

Urban spatio-temporal (ST) forecasting is crucial for various urban applications such as intelligent scheduling and trip planning. Previous studies focus on modeling ST correlations among urban locations in offline settings, which often neglect the non-stationary nature of urban ST data, particularly, distribution shifts over time. This oversight can lead to degraded performance in real-world scenarios. In this paper, we first analyze the distribution shifts in urban ST data, and then introduce DOST, a novel online continual learning framework tailored for ST data characteristics. DOST employs an adaptive ST network equipped with a variable-independent adapter to address the unique distribution shifts at each urban location dynamically. Further, to accommodate the gradual nature of these shifts, we also develop an awake-hibernate learning strategy that intermittently fine-tunes the adapter during the online phase to reduce computational overhead. This strategy integrates a streaming memory update mechanism designed for urban ST sequential data, enabling effective network adaptation to new patterns while preventing catastrophic forgetting. Experimental results confirm DOST's superiority over state-of-the-art models on four real-world datasets, providing online forecasts within an average of 0.1 seconds and achieving a 12.89% reduction in forecast errors compared to baseline models.

Figures (7)

• Figure 1: An illustration of Chicago's taxi demand distribution shift, estimated with Kernel Density Estimator (KDE). Left hand side: estimated distributions for Region 23 (upper) and Region 64 (lower); right hand side: a visualization of Chicago's community regions.
• Figure 2: Overview of DOST, which employs two strategies: (a) Adaptive spatio-temporal (ST) network for online learning, where modules are represented in three colors: gray for traditional modules, green for the adapter, and yellow for the awake decider. (b) Awake-Hibernate (AH) learning strategy, which alternates network updates between awake and hibernate phases. During the awake phase, the adapter is fine-tuned using the Streaming Memory Update (SMU) mechanism, while during the hibernate phase, all parameters are frozen. Note: The Memory Placeholder in the SMU mechanism is omitted in this figure.
• Figure 3: Overview of Variable-Independent Adapter (VIA), which comprises a series of Sub-VIAs, each dedicated to learning the distinct distribution shifts at specific urban locations.
• Figure 4: An illustration of the Streaming Memory Update (SMU) Mechanism.
• Figure 5: Online Urban Spatio-Temporal Forecasting