KG4RSV: Towards Knowledge Graph -driven Specification and Validation of ML Dataset Quality Requirements: Case Study on e -ICU dataset
Year: 2026
Type: article
Abstract: Artificial Intelligence (AI) and Machine Learning (ML) are increasingly embedded in critical domains, raising the challenges for ensuring the quality and reliability of AI/ML solutions. While traditional Requirements Engineering (RE) plays a central role in eliciting and specifying the stockholders’ expectations, current practices often lack the systematic specification and verification of dataset requirements, even though poor data quality issues remain a primary cause of ML project failures. This paper introduces a Knowledge Graph-enabled framework for ML Dataset Requirements Specification (KG4RSV) and Validation. KG4RSV addresses three critical challenges in dataset engineering: managing data heterogeneity, ensuring data quality, and supporting systematic specification and validation of dataset requirements. Leveraging the semantic capabilities of Knowledge Graphs (KGs), the proposed method enables a structured representation of both data requirements and datasets using Resource Description Framework (RDF) and SHACL standards. We evaluate KG4RSV on a real-world clinical dataset electronic Intensive Care Unit (eICU) for binary classification of cardiovascular events. Results demonstrate that the framework effectively detects and handles quality issues related to completeness, consistency, and semantic correctness. The validation process led to notable improvements in ML performance-up to a 7% increase in recall for an Artificial Neural Network (ANN) model and an average 5% improvement across multiple ML models, such as Random Forest (RL) and Logistic Regression (LR). By bringing dataset validation into the requirements process, KG4RSV improves data quality and ensures better alignment between data and system goals from the early phases.
Author: Ziad NASRI
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7214684635