Revisiting Node Affinity Prediction in Temporal Graphs

2025-10-08

AI Summary

Existing temporal graph neural networks (TGNs) underperform simple heuristic baselines—such as Persistent Forecast and Moving Average—on node affinity prediction, revealing fundamental flaws in their training mechanisms. This work systematically diagnoses the issue, identifying gradient instability and state modeling mismatch as primary causes. To address these, we propose NAViS: a novel TGN architecture featuring a virtual state mechanism that establishes, for the first time, theoretical equivalence between heuristic predictors and state-space models; and a gradient-aware adaptive loss function that mitigates training bias. Evaluated on the Temporal Graph Benchmark (TGB), NAViS achieves state-of-the-art performance—outperforming all existing TGNs and heuristic baselines—while offering improved interpretability and trainability. Our approach introduces a principled, optimization-friendly paradigm for dynamic graph modeling.

Technology Category

• Machine Learning: Graph-based Machine Learning
• Planning, Routing, and Scheduling: Temporal Planning
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Social Networks and Social Media: Influence propagation, information diffusion, and the prediction on networks
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Node affinity prediction is a common task that is widely used in temporal graph learning with applications in social and financial networks, recommender systems, and more. Recent works have addressed this task by adapting state-of-the-art dynamic link property prediction models to node affinity prediction. However, simple heuristics, such as Persistent Forecast or Moving Average, outperform these models. In this work, we analyze the challenges in training current Temporal Graph Neural Networks for node affinity prediction and suggest appropriate solutions. Combining the solutions, we develop NAViS - Node Affinity prediction model using Virtual State, by exploiting the equivalence between heuristics and state space models. While promising, training NAViS is non-trivial. Therefore, we further introduce a novel loss function for node affinity prediction. We evaluate NAViS on TGB and show that it outperforms the state-of-the-art, including heuristics. Our source code is available at https://github.com/orfeld415/NAVIS

Problem

• Analyzing training challenges in Temporal Graph Neural Networks
• Developing node affinity prediction model using virtual states
• Introducing novel loss function for improved prediction accuracy

Innovation

• Uses virtual state to model node affinity
• Exploits equivalence between heuristics and state space models
• Introduces novel loss function for training optimization