2. Our participation — and motivation — in the
project involved the research &
development of a recommendation engine
that...
● leveraged the ubiquitousness and richness of
linked data from the Web of Data
● would produce new linked data as a result of
those recommendations. In addition, this would
provide data interlinking
3. In general, we were concerned with...
●How to perform recommendation computations
with the linked data? Furthermore, how to do this
scalably?
● How to input linked data into such a system?
●How to output linked data from those
recommendations?
4. For recommendation types, we focused on
implementing the primary types:
Collaborative filtering &
Content-based
With an array of algorithms including — Cosine Similarity, Pearson
Correlation, Jaccard Distance, Co-occurrence, etc.
●An initial direction for the computation of
recommendations ✓
Challenge: adapting these algorithms for linked
●
data ✓
5. SPARQL for the Input of Linked Data
SELECT ?s ?nationality ?influences WHERE {
?s dbpedia-ontology:occupation dbpedia-
resource:Poet.
?s dbpedia-property:influences ?influences.
?s dbpedia-ontology:nationality ?nationality.
}
●Declarative and expressive method for data materialisation ✓
●SPARQL endpoint communication ✓
7. Sometimes...
Linked data → Big data
Therefore, we went in the direction of a distributed and
parallel framework — MapReduce
8. Overview of Results
● SPARQL execution, RDF materialisation and
output → design the system using established
tools and libraries
● The adaptation of the recommendation
algorithms for RDF → formalisations
presented in a paper [ECAI 2012]
● Scalability with the possibly large amount of
data that can be input → a parallel and
distributed framework
10. Deployment and Use
To get SLDR running, a JSP web server, such
as Tomcat or Jetty is required.
SLDR is deployed as a web application (WAR).
From there, the recommendation engine is
fully accessible from your web browser to start
creating and running jobs.
14. Retrieving Recommendations
Users have the option of viewing computed
recommendations through either SPARQL and
the output triplestore or through a REST API
implemented into the systems backend.
The REST API can be utilised for better
integration into already existing systems (e.g.
HTML, JavaScript, etc.)
15. Summary
● Ongoing improvement and development
● Have tested sucessfully with some of the
SME's
● More information available at
http://sldr.deri.ie