Uncertainty-Regularized Neural Surrogates for Sparse Governing- Drift-Aware Correspondence Learning Between Structurally Mismatched Temporal Knowledge Graphs

Abstract

A B S T R A C T

Temporal knowledge graphs have emerged as a critical structural representation for capturing dynamic relational data across various domains, including finance, healthcare, and global event monitoring. As organizations construct independent knowledge bases, the need to integrate these disparate sources through entity alignment or correspondence learning becomes paramount. However, integrating temporal knowledge graphs presents unique challenges that extend beyond static graph alignment. Specifically, real-world temporal knowledge graphs frequently suffer from structural mismatch due to divergent schema designs and data collection methodologies. Furthermore, they are subject to semantic drift, a phenomenon where the relational context and fundamental characteristics of entities evolve asynchronously over time. This paper proposes a comprehensive framework for drift-aware correspondence learning designed specifically to address the intersection of structural mismatch and temporal semantic drift. By utilizing a novel text-descriptive temporal encoding strategy combined with a structural reconciliation module, the proposed approach maps heterogeneous graph structures into a unified temporal latent space. The framework incorporates a drift-tracking mechanism that quantifies the temporal evolution of entity representations, ensuring that alignment decisions remain robust even when entity behaviors diverge. Through extensive conceptual analysis and systematic evaluation on adapted real-world datasets, the proposed methodology demonstrates substantial improvements over traditional static and time-aware alignment techniques. The findings underscore the necessity of modeling temporal dynamics and structural heterogeneity concurrently, paving the way for more resilient knowledge integration systems in dynamically evolving environments.