Libraly Journal

Libraly Journal ›› 2026, Vol. 45 ›› Issue (10): 100-115.

Previous Articles     Next Articles

Research on Scientific Dataset Profiling Based on Knowledge Graphs

Wu Jingya, Cai Hongyu, Shi Jin   

  • Online:2026-10-15 Published:2026-10-27
  • About author:Wu Jingya, Cai Hongyu, Shi Jin

Abstract: Driven by open science and AI4S, scientific datasets have become core resources for knowledge generation and research innovation. However, existing methods for describing and evaluating datasets are limited to surface-level attributes, failing to support efficient dataset discovery and value assessment. Based on the concept of data profiling, this paper proposes the concept of dataset profiling to explore a systematic approach to characterizing scientific datasets and break through the current bottleneck in dataset description. A knowledge graph-based profiling method is proposed, constructing a multi-dimensional labeling system covering basic attributes, relational characteristics, and impact, and realizing structured modeling and semantic expression via knowledge graphs. Taking 6603 machine learning datasets from the papers with code platform as the research sample, the study validates the approach through label extraction, knowledge graph construction, hierarchical clustering  analysis. A hierarchical structure of three dataset types is successfully identified: core benchmark datasets, mainstream application datasets, and peripheral long-tail datasets. The proposed approach effectively expands the theoretical and methodological system of scientific dataset research, and provides new tools and pathways for dataset discovery and adaptation assessment in data-driven research.

Key words: Dataset&emsp, description,&emsp, Data&emsp, profiling,&emsp, Dataset&emsp, profiling,&emsp, Knowledge&emsp, graph,&emsp, Scientific&emsp, dataset