Predictive insights transfer between different temporal graph types
Learning Across Temporal Graph Representation Silos
Machine LearningArtificial Intelligence
Summary
Graphs that change over time can be shown in different ways, but it’s unclear if models trained on one kind can help with another. The authors found that predictive patterns learned from one temporal graph form can be reused in a different type, even when these forms use separate vocabularies and interfaces. This reuse improves predictions and works without needing to align the entities or relations exactly. They demonstrated this idea with examples like recommendation systems and security detection where different data sources helped each other.
What this means in practice
- •For recommendation system developers: Use pretrained models from different temporal graph types to improve recommendation accuracy without needing aligned data.
- •For security monitoring teams: Combine temporal graph models from diverse sources to enhance anomaly detection without direct entity or relation matching.
Authors
Minwoo Yu, Young-guk Ha
Abstract
Temporal graph data are recorded in different representations, but whether these differences also determine the scope of learned knowledge reuse remains unclear. This study investigates whether predictive structure learned from relational histories can be reused between continuous-time dynamic graphs (CTDGs) and temporal knowledge graphs (TKGs). Bidirectional reuse emerges while retaining distinct input and prediction interfaces and independent vocabularies, and sources with the same representation are not always more useful. Controlled reorganizations of source histories show that the organization contributing to reuse varies across source-target conditions. We show that learned aspects of Relational Dynamics, the predictive structure present in relational histories, can be reused across conventional representation boundaries. Independently pretrained TKG and CTDG learners can also be fused at the score level to improve prediction on a third dataset without entity or relation-semantic alignment. Two proofs of concept in recommendation and security detection illustrate that external temporal sources without direct alignment can also become useful pretraining resources.