Efficient and Extensive Construction of Knowledge Graphs from Diverse Data with Declarative Mappings
- Autores: Julián Arenas Guerrero
- Directores de la Tesis: Oscar Corcho García (dir. tes.), María de los Santos Pérez Hernández (codir. tes.)
- Lectura: En la Universidad Politécnica de Madrid ( España ) en 2025
- Idioma: inglés
- Tribunal Calificador de la Tesis: José Emilio Labra Gayo (presid.), Andrea Jesús Cimmino Arriaga (secret.), Paola Patricia Espinoza Arias (voc.), Eduardo Mena Nieto (voc.), Sebastián Ferrada Aliaga (voc.)
- Programa de doctorado: Programa de Doctorado en Inteligencia Artificial por la Universidad Politécnica de Madrid
- Calificación: Sobresaliente Cum Laude
- Materias:
- Enlaces
- Tesis en acceso abierto en: Archivo Digital UPM
- Resumen
Heterogeneous data is typically integrated into a knowledge graph using declarative mappings. There are two ways of constructing the graph with the mappings: materialization and virtualization. Historically, virtualization has focused on relational databases due to their predominance and because of the possibility of pushing down costly computations in SQL. However, the current prevalence of diverse data has rendered existing virtualization techniques insufficient. This diversity is not a problem in materialization, but it must be optimized to create large graph dumps. The first research problem addressed in this thesis is the optimization of materialization. We propose mapping partitioning, which divides rules into groups with non-overlapping outputs. Two strategies are introduced to optimize materialization: parallel processing of the mapping groups reduces execution times, and sequential processing reduces the consumption of memory. We devise algorithms for partitioning a mapping and validate their effectiveness on several benchmarks and use cases. We empirically compare materialization with mapping partitioning against a baseline without partitioning and state-of-the-art systems. The results demonstrate the positive impact of partitioning on materialization with a large reduction in execution times and memory usage. Building on optimized materialization, we address our second research problem: generic virtualization over multiple and diverse sources. We propose shifting from the established query translation approach to a data translation approach. The idea is to converge on the fly the relevant subset of data for a SPARQL query into an intermediate representation. Coalescing heterogeneous data into this representation is easier than translating queries into the potentially diverse query languages of the local sources. We present the ITT architecture, based on star-shaped mapping candidate selection to find the relevant set of rules, rapid materialization of intermediate results into a triple table, and unfolding SPARQL queries to SQL over the triple table. Experiments with mixed data sources demonstrate the generic nature of ITT. Compared to query translation with Ontop, the experimental results showed that ITT outperformed query translation in many cases. While ITT exhibited a more intensive use of memory, it did not experience timeouts, unlike Ontop. The thesis also addresses other challenges in KG construction arising from the emergence of the RDF-star graph model and the need for flexible data transformations. For RDF-star, we devise the RML-star mapping language, a baseline materialization algorithm, and a method for the automatic extraction of RML-star from property graphs to mitigate labor-intensive mapping development. Regarding flexible data transformations, we extend RML views to tabular files to leverage SQL, overcoming RML's limitations with complex joins and composite data values. We also propose unfolding data transformations defined in RML-FNML to RML views, enabling query translation-based virtualization systems to process RML-FNML with prior unfolding of the mappings. Finally, we implemented Python user-defined functions in RML to allow more flexible data transformations using a high-level programming language; their positive impact on knowledge graph construction has been demonstrated at an industry scale in the chemical sector.
