An Effective Embedding Approach to Shortest Path Distance Prediction Over Large-Scale Graphs

ICLR 2026 Conference Submission15609 Authors

Keywords: Shortest Path Distance, Graph Algorithm, Representation Learning, Neural Network Model

Abstract:

In graph data management, computing the shortest path distance between any pair of nodes is a crucial and foundational graph operation with numerous practical applications (e.g., travel/route planning, community search).
Traditional algorithms for solving this problem face significant challenges in time and space complexities, especially when dealing with large-scale graphs. Worse still, existing learning-based approaches often struggle with low accuracy in predicting intricate graph structures.
To address these issues, this paper introduces a novel Graph Convolutional Networks (GCN)- and Multi-View Deep Neural Networks (MVDNN)-based Distance Embedding (GM-DE) framework, which enables fast and accurate predictions of the shortest path distances. Specifically, based on our proposed pivot and anchor set selection strategies,
GM-DE enables the calculation of embeddings for each graph node.
Then, by feeding such embeddings into our designed GCN and MVDNN models, GM-DE can be well trained to support the mining of accurate global and local positional information for the graph nodes with the help of our constructed predictors.
In this way, our GM-DE framework can achieve high accuracy in various complex scenarios, relying solely on basic node attributes as input without the need for scenario-specific data.
Comprehensive experiments confirm the effectiveness and efficiency of GM-DE approach in predicting the shortest path distances on a wide range of real-world graphs.

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

Submission Number: 15609