Temporal Link Prediction Using Graph Embedding Dynamics

TL;DR

This work reframes temporal link prediction by treating nodes as Newtonian objects with position and velocity in a learned embedding space. A velocity predictor (3-layer LSTM) projects future node velocity, which, through vertical and horizontal aggregations, yields predicted node locations used to assess future links via Euclidean similarity. The approach enhances both predictive accuracy and interpretability over traditional temporal embeddings, achieving a 17.34% AUROC improvement on long-running co-authorship data and outperforming several state-of-the-art methods on benchmark datasets. The method is generalizable to other dynamic graphs and requires only modest historical history, enabling efficient multi-step forecasting and broader application beyond co-authorship networks.

Abstract

Graphs are a powerful representation tool in machine learning applications, with link prediction being a key task in graph learning. Temporal link prediction in dynamic networks is of particular interest due to its potential for solving complex scientific and real-world problems. Traditional approaches to temporal link prediction have focused on finding the aggregation of dynamics of the network as a unified output. In this study, we propose a novel perspective on temporal link prediction by defining nodes as Newtonian objects and incorporating the concept of velocity to predict network dynamics. By computing more specific dynamics of each node, rather than overall dynamics, we improve both accuracy and explainability in predicting future connections. We demonstrate the effectiveness of our approach using two datasets, including 17 years of co-authorship data from PubMed. Experimental results show that our temporal graph embedding dynamics approach improves downstream classification models' ability to predict future collaboration efficacy in co-authorship networks by 17.34% (AUROC improvement relative to the baseline model). Furthermore, our approach offers an interpretable layer over traditional approaches to address the temporal link prediction problem.

Figures (6)

• Figure 1: Velocity prediction, followed by link prediction. (Top) Tracking the movement of nodes in the embedding space from their location in the first year moving to the second year. ( Middle) Predicting the future location of nodes in the embedding space based on their current location (second/last year, shown by squares) and the movement's dynamics; i.e., velocity. (Bottom) Predicting links between the predicted location of nodes in the future based on their similarity in the embedding space.
• Figure 2: Performance Evaluation of Different Embedding Sizes in terms of AUROC Score
• Figure 3: Performance Evaluation of Different Embedding Sizes in terms of AUPRC Score
• Figure 4: Performance Evaluation of Different Time-series Length in terms of AUROC Score
• Figure 5: Performance Evaluation of Different Time-series Length in terms of AUPRC Score