Bridging Temporal and Semantic Gaps: Prompt Learning on Temporal Interaction Graphs

ICLR 2026 Conference Submission5531 Authors

Keywords: Dynamic Graph, Temporal Interaction Graph, Graph Prompt Learning

Abstract:

Temporal Interaction Graphs (TIGs) are widely utilized to represent real-world systems like e-commerce and social networks. While various TIG models have been proposed for representation learning, they face two critical gaps in their "pre-train, predict" training paradigm: a temporal gap limiting timely predictions and a semantic gap reducing adaptability to diverse downstream tasks.
A potential solution is applying the ``pre-train, prompt'' paradigm, yet existing static graph prompting methods fail to address the time-sensitive dynamics of TIGs and have a deficiency in expressive power.
To tackle these issues, we propose **Temporal Interaction Graph Prompting (TIGPrompt)**, a versatile framework that bridges the temporal and semantic gaps by integrating with existing TIG models. Specifically, we propose a "pre-train, prompt" training paradigm for TIGs, with a temporal prompt generator to offer temporally-aware prompts for different tasks. To cater to varying computational resource demands, we propose an extended "pre-train, prompt-based fine-tune" paradigm, offering greater flexibility.
Through extensive experiments involving multiple benchmarks, representative TIG models, and downstream tasks, our TIGPrompt demonstrates the SOTA performance and remarkable efficiency advantages. The codes are available at an [Anonymous Repository](https://anonymous.4open.science/r/TIGPrompt).

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 5531