The Data Web has been growing significantly over the last decades. Thus, linking resources across heterogeneous knowledge bases becomes an increasingly difficult problem, in particular w.r.t. the runtime of link discovery tasks. Consequently, it is of utmost importance to provide time-efficient approaches for link discovery on the Data Web. While a number of scalable approaches have been developed for various measures, the most frequent k characters measure has not been tackled in previous works. We hence present a sequence of filters that allow discarding comparisons when executing bounded similarity computations without losing any recall. Therewith, we can reduce the runtime of bounded similarity computations by approximately 70%. Our experiments with a single-threaded, a parallel and a GPU implementation of our filters suggests that our approach scales well even when dealing with millions of potential comparisons.
- Java 8+ (http://www.oracle.com/technetwork/java/javase/downloads/jdk8-downloads-2133151.html)
- Jena API (https://jena.apache.org/download/index.cgi).
Run the main method in the following classes to see the results.
- MFKC.java is the implementation of the algorithm.
- MFKC_Parallel.java is the parallel implementation using Java 8 resources.
- In MFKC_Sep.java, the filters are executed separately.
- ExperimentMFKC.java contains all experiments.
Datasets used:
- dbPediaPlaces.tsv is the dataset from DBpedia
- linkedGeoPlaces_million.tsv is the dataset from LinkedGeoData
- dbpedia_yago_labels_10k.tsv contains label alignments - download
This work was published as a full paper was published at ISWC 2017 on the Ontology Matching workshop.
http://disi.unitn.it/~pavel/om2017/papers/om2017_Tpaper1.pdf