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.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve the problem of distribution shifts in urban spatio - temporal (ST) data in online continual learning. Specifically, the paper focuses on: 1. **Non - stationarity problem**: Previous studies mainly focused on modeling spatio - temporal correlations between different locations in the city in an offline environment, while ignoring the non - stationary characteristics of urban ST data, especially the distribution shifts in time. This neglect will lead to performance degradation in practical application scenarios. 2. **Dynamics of distribution shifts**: The distribution of urban ST data changes gradually over time, not just suddenly. For example, the data distribution in some areas may remain stable in a short time, but change significantly over a longer time span (such as the Chicago taxi demand distribution shown in Figure 1). Traditional instant - update or batch - update methods are not applicable in this case. 3. **Location - specific distribution shifts**: The distribution shift patterns in different urban locations are different. For example, the data distribution in school areas may change drastically during holidays, while the data distribution in the central business district (CBD) is relatively stable. Therefore, a method that can adapt to location - specific distribution shifts is needed. To solve these problems, the authors propose a new framework named DOST (Distribution - aware Online Continual Learning for Urban Spatio - Temporal Forecasting). DOST addresses the above challenges in the following ways: - **Adaptive spatio - temporal network**: An adaptive spatio - temporal network with a Variable - Independent Adapter (VIA) is introduced to dynamically handle the unique distribution shifts in each urban location. - **Awake - Hibernate learning strategy**: An Awake - Hibernate (AH) learning strategy is developed. This strategy intermittently fine - tunes the adapter to reduce computational overhead and effectively adapt to new patterns through the Streaming Memory Update (SMU) mechanism, preventing catastrophic forgetting. Through these methods, DOST can provide more accurate online predictions in real - time forecasting and achieves higher performance than existing models on four real - world datasets, with an average prediction error reduction of 12.89%. ### Summary The core problem of the paper is to solve the distribution shift problem in urban spatio - temporal data in online continual learning, especially the challenges in non - stationarity and location - specific distribution shifts. The DOST framework effectively improves the accuracy and efficiency of urban spatio - temporal forecasting by introducing an adaptive spatio - temporal network and an Awake - Hibernate learning strategy.

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

Space-Time Adaptive Network for Origin-Destination Passenger Demand Prediction

Boosting Urban Prediction via Addressing Spatial-Temporal Distribution Shift

DynST: Dynamic Sparse Training for Resource-Constrained Spatio-Temporal Forecasting

Dynamic Spatiotemporal Straight-Flow Network for Efficient Learning and Accurate Forecasting in Traffic

ODEStream: A Buffer-Free Online Learning Framework with ODE-based Adaptor for Streaming Time Series Forecasting

HSTFL: A Heterogeneous Federated Learning Framework for Misaligned Spatiotemporal Forecasting

Dual-track Spatio-temporal Learning for Urban Flow Prediction with Adaptive Normalization

Spatial-temporal Forecasting for Regions without Observations

Online Spatio-Temporal Crowd Flow Distribution Prediction for Complex Metro System

DeepSTCL: A Deep Spatio-temporal ConvLSTM for Travel Demand Prediction

Memory-enhanced Invariant Prompt Learning for Urban Flow Prediction under Distribution Shifts

EasyST: A Simple Framework for Spatio-Temporal Prediction

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

Maintaining the Status Quo: Capturing Invariant Relations for OOD Spatiotemporal Learning

FairSTG: Countering performance heterogeneity via collaborative sample-level optimization

Deep Learning for Metro Short-Term Origin-Destination Passenger Flow Forecasting Considering Section Capacity Utilization Ratio

Learning Graph Structures and Uncertainty for Accurate and Calibrated Time-series Forecasting

DNN-based Prediction Model for Spatio-Temporal Data

STTM: A New Approach Based Spatial-Temporal Transformer And Memory Network For Real-time Pressure Signal In On-demand Food Delivery

Combating Distribution Shift for Accurate Time Series Forecasting via Hypernetworks