[Paper Review] MapSDI: A Scaled-up Semantic Data Integration Framework for Knowledge Graph Creation

Samaneh Jozashoori, María-Esther Vidal|arXiv (Cornell University)|Sep 3, 2019

Semantic Web and Ontologies13 references17 citations

TL;DR

MapSDI is a scalable semantic data integration framework that optimizes knowledge graph creation by pre-processing heterogeneous, large-scale datasets using mapping rules (e.g., RML) to project relevant attributes and eliminate duplicates before RDFization. By leveraging relational algebra properties and semantic knowledge from mapping rules, MapSDI reduces RDFizer execution time by up to one order of magnitude, with no data loss, significantly improving performance on big data with high volume, variety, and veracity.

ABSTRACT

Semantic web technologies have significantly contributed with effective solutions for the problems of data integration and knowledge graph creation. However, with the rapid growth of big data in diverse domains, different interoperability issues still demand to be addressed, being scalability one of the main challenges. In this paper, we address the problem of knowledge graph creation at scale and provide MapSDI, a mapping rule-based framework for optimizing semantic data integration into knowledge graphs. MapSDI allows for the semantic enrichment of large-sized, heterogeneous, and potentially low-quality data efficiently. The input of MapSDI is a set of data sources and mapping rules being generated by a mapping language such as RML. First, MapSDI pre-processes the sources based on semantic information extracted from mapping rules, by performing basic database operators; it projects out required attributes, eliminates duplicates, and selects relevant entries. All these operators are defined based on the knowledge encoded by the mapping rules which will be then used by the semantification engine (or RDFizer) to produce a knowledge graph. We have empirically studied the impact of MapSDI on existing RDFizers, and observed that knowledge graph creation time can be reduced on average in one order of magnitude. It is also shown, theoretically, that the sources and rules transformations provided by MapSDI are data-lossless.

Motivation & Objective

• To address scalability challenges in semantic data integration for knowledge graph creation, especially with big data characterized by high volume, variety, and veracity.
• To reduce the performance overhead caused by redundant data and inefficient processing in existing RDFizers like RMLmapper and SDM-RDFizer.
• To enable efficient, lossless pre-processing of input datasets by leveraging semantic knowledge encoded in mapping rules.
• To integrate schema-level and data-level integration into a single optimized pipeline, avoiding costly post-processing steps.

Proposed method

• MapSDI uses mapping rules (e.g., RML) to identify required attributes and apply semantic-aware pre-processing to input datasets.
• It applies relational algebra operations—projection and selection—based on rule semantics to reduce data size before RDFization.
• The framework pushes down projections and selections into the data processing pipeline to minimize the input volume for the RDFizer.
• It performs duplicate elimination based on attribute sets defined in mapping rules, reducing data redundancy early in the pipeline.
• MapSDI ensures data-lossless transformations by preserving all semantic information required for correct RDF triple generation.
• The framework is evaluated with real-world biomedical datasets and integrated with existing RDFizers (RMLmapper, SDM-RDFizer) to measure performance gains.

Experimental results

Research questions

• RQ1 Can pre-processing of input data using semantic knowledge from mapping rules significantly reduce RDFizer execution time in knowledge graph creation?
• RQ2 How does MapSDI’s data-lossless pre-processing affect the scalability of existing RDFizers on large, heterogeneous, and low-quality datasets?
• RQ3 To what extent do attribute projection and duplicate removal improve performance when joining multiple datasets via RML triple maps?
• RQ4 Can MapSDI be applied independently of the underlying mapping language or RDFizer, and does it maintain consistent performance gains across different tools?

Key findings

• MapSDI reduces RDFizer execution time by up to one order of magnitude (on average) across multiple test datasets.
• The framework achieves a 98% reduction in input dataset size (from 235,000 KB to 997 KB) after pre-processing on a 100% duplicate dataset, significantly improving performance.
• Even with high duplicate rates (up to 100%), MapSDI maintains stable performance, while baseline frameworks (T-framework) show increasing execution time with data volume and duplication.
• MapSDI enables the SDM-RDFizer to complete joins that previously timed out (500 seconds), demonstrating improved scalability under complex mapping conditions.
• Theoretical analysis confirms that MapSDI’s pre-processing steps are data-lossless, ensuring the final knowledge graph remains semantically identical to the original integration pipeline.
• Empirical results show that MapSDI outperforms baseline systems regardless of the RDFizer used, with consistent performance gains across varying data sizes and duplicate ratios.