Semantic-Aware Graph Schema Discovery for Heterogeneous Cloud Metadata
Year: 2026
Type: conference-paper
Abstract: Effective governance of large-scale multi-cloud infrastructures critically depends on consistent and interpretable resource metadata. However, heterogeneous teams and tools induce severe semantic drift in cloud tags, preventing automated schema enforcement and policy validation. This paper formalizes cloud metadata as unordered semantic property sets within a property graph, demonstrating that existing schema discovery fails under realistic tagging inconsistencies. We also show that sequential Large Language Model (LLM) embeddings, although semantically expressive, introduce permutation-sensitive variance incompatible with flat metadata sets. To address this gap, we design a permutation-invariant contrastive representation learning pipeline tailored to the semi-structured nature of cloud metadata. By combining LLM-based encodings with set-based attention, our approach enables stable clustering and hierarchical schema reconstruction in highly heterogeneous metadata environments. Experiments on a real-world, multi-tenant production infrastructure demonstrate high clustering stability despite severe metadata noise. Crucially, our approach yields superior geometric cluster separation and graceful degradation under noise compared to baselines. Ultimately, this provides a practical framework for automated cloud governance and FinOps optimization.
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204446565