Work

A drift-adaptive learned index with knowledge-graph-enhanced hybrid retrieval improves metadata management for large-scale scientific computing data

Year: 2026
Type: article
Abstract: Metadata from large scale scientific computing grows rapidly, and its high dimensional heterogeneity, continuous evolution and domain semantics strain both traditional index structures and keyword based retrieval, which have developed in isolation. This paper proposes MetaSciX, a metadata management and hybrid semantic retrieval framework built on four techniques: a multidimensional learned index that combines adaptive quantile quantization with a z-order curve and absorbs distribution drift by statistically triggered incremental fine-tuning; a knowledge graph enhanced pipeline that encodes domain relations with a graph attention network and fuses sparse, dense and graph signals under query aware weights; consistent hashing with virtual nodes for distributed consistency; and a version aware provenance scheme that folds record versions into the composite key and keeps PROV compatible lineage in the knowledge graph. On real high energy physics and climate metadata, MetaSciX lowers point query latency by at least 30.5% relative to the PGM-Index, uses at most 15.4% of the K-D tree memory, and raises Recall@10 by at least 8.1% over the best single channel retriever and 4.1% over static fusion. On 100 M synthetic records, 8 nodes reach 12,420 queries per second at a scaling efficiency of 0.84, and ablation confirms the contribution of each component.
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7236630148
FWCI:
Cites: 37
Cited by:
Related to: 10