LOD Lab: Experiments at LOD Scale

Contemporary Semantic Web research is in the business of optimizing algorithms for only a handful of datasets such as DBpedia, BSBM, DBLP and only a few more. This means that current practice does not generally take the true variety of Linked Data into account. With hundreds of thousands of datasets out in the world today the results of Semantic Web evaluations are less generalizable than they should and — this paper argues — can be. This paper describes LOD Lab: a fundamentally different evaluation paradigm that makes algorithmic evaluation against hundreds of thousands of datasets the new norm. LOD Lab is implemented in terms of the existing LOD Laundromat architecture combined with the new open-source programming interface Frank that supports Web-scale evaluations to be run from the command-line. We illustrate the viability of the LOD Lab approach by rerunning experiments from three recent Semantic Web research publications and expect it will contribute to improving the quality and reproducibility of experimental work in the Semantic Web community. We show that simply rerunning existing experiments within this new evaluation paradigm brings up interesting research questions as to how algorithmic performance relates to (structural) properties of the data.