Prompt-Based Spatio-Temporal Graph Transfer Learning

TL;DR

This work proposes Spatio-Temporal Graph Prompting (STGP), a prompt-based framework capable of adapting to multi-diverse tasks in a data-scarce domain and employs learnable prompts to achieve domain and task transfer in a two-stage prompting pipeline, facilitating the prompts to effectively capture domain knowledge and task-specific properties.

Abstract

Spatio-temporal graph neural networks have proven efficacy in capturing complex dependencies for urban computing tasks such as forecasting and kriging. Yet, their performance is constrained by the reliance on extensive data for training on a specific task, thereby limiting their adaptability to new urban domains with varied task demands. Although transfer learning has been proposed to remedy this problem by leveraging knowledge across domains, the cross-task generalization still remains under-explored in spatio-temporal graph transfer learning due to the lack of a unified framework. To bridge the gap, we propose Spatio-Temporal Graph Prompting (STGP), a prompt-based framework capable of adapting to multi-diverse tasks in a data-scarce domain. Specifically, we first unify different tasks into a single template and introduce a task-agnostic network architecture that aligns with this template. This approach enables capturing dependencies shared across tasks. Furthermore, we employ learnable prompts to achieve domain and task transfer in a two-stage prompting pipeline, facilitating the prompts to effectively capture domain knowledge and task-specific properties. Our extensive experiments demonstrate that STGP outperforms state-of-the-art baselines in three tasks-forecasting, kriging, and extrapolation-achieving an improvement of up to 10.7%.

Figures (6)

• Figure 1: Comparison of existing transfer learning approaches with our prompt tuning. (a) The pre-trained model is adapted to a target domain, aided by a domain transfer model. (b) The spatio-temporal model is pre-trained on a data-rich source domain. (c) STGP leverages prompts to modify data, achieving domain and task transfer, respectively.
• Figure 2: Spatio-temporal graph downstream tasks.
• Figure 3: The overview of our Spatio-Temporal Graph Prompting framework. (a) The network architecture is task-agnostic and aligns with the unified template. (b) Domain and task prompts are learned sequentially through the two-stage prompting pipeline. (c) The tasks are consolidated within a data reconstruction template, where they are represented by masked locations.
• Figure 4: Model analysis with METR-LA as the target domain.
• Figure 5: Effect of prompt numbers and the number of target days for training on the METR-LA dataset.