The availability of a data source as huge, open, accessible and machine-understandable as the Web of Data asks for new and sophisticated inferences to be implemented in order to deeply exploit such a rich informative content. Towards this direction, the paper proposes an approach for inferring clusters in collections of RDF resources on the basis of the features shared by their descriptions. The approach grounds on an algorithm for Common Subsumers computation proposed in a previous work of some of the authors. The clustering service introduced here returns not only di↵erent cluster proposals for a given collection, but also a description of the informative content shared by the RDF resources within the clusters, in terms of (generalized) RDF triples.
Introduction a knowledge domain. In particular, we chose LOD by Chamber of Deputies of Italian Parliament5 as case study.
The paper is organized as follows: in the next section we describe the main features of the proposed approach, together with some details on its implementation. In Section 3, some preliminary results are shown, before closing the paper. 2
The Approach The approach we propose here aims at automatically clustering a target collection of RDF resources according to a fully semantic-based classification. In particular, RDF descriptions are investigated to infer clusters of resources entailing the same sets of RDF triples, in order to provide a description of the informative content shared within each cluster.
The originality of the proposal lays in the choice of adopting deductive
services to learn6 clusters description from examples represented in RDF. In fact,
although clustering is a thoroughly investigated task in machine learning
literature, approaches solving it usually adopt induction to identify clusters according
to some—sometimes semantic-based—distance between elements in the same
cluster ([
We propose to solve clustering through a deductive and fully semantic-based approach, which relies on the iteration of the following two steps: i) the CS of two randomly selected RDF resources (in the following referred as seed-resources) is computed; ii) the rest of the target collection is queried in order to find other items entailing the same CS. The sub-collection made up by the two initial resources and those returned by step ii) is one cluster of the collection. Therefore, it is subtracted from the initial target collection, and the two steps are iterated until there are no more resources to be clustered.
An anytime algorithm to compute a CS of pairs of RDF resources has been
proposed in [
Then, the set T of triples is used to model a SPARQL [
In order to exemplify our clustering approach, we adopt the LOD by
Chamber of Deputies of Italian Parliament as use case. Such an informative source
5 http://dati.camera.it/data/en/
6 In Machine Learning vocabulary, what we do is called unsupervised learning
7 Due to space limits, the CS extraction algorithm [
In Figure 1, the reader may find two example r-graphs describing a pair of
resources (ocd:d3140 10 and ocd:d270 10) in our reference dataset. In the most
general case, all triples of datasets of interest having the seed-resource as subject
are considered relevant for the description of the resource itself. Here, we adopt a
more restrictive strategy driven by specific knowledge of the domain, and discard
as not relevant for the description of a resource r also triples <<r p o>> such
that p 2 { dc:date, dc:title, foaf:depiction, foaf:firstName, foaf:nick,
foaf:surname, ocd:endDate, ocd:file, ocd:startDate, ods:modified, rdfs:
comment, rdfs:label, terms:isReferencedBy}.
In Table 1, we report a clustering proposal for all deputies of the 10th legislature,
where resources ocd:d3140 10 and ocd:d270 10 of the above example10 were
forcibly selected as seed pair for the extraction of the first cluster (P1). By
looking at the first row, one can notice how no other resource in the collection
shares the features of the CS originated by the first seed-pair, i.e., |P1| = 2. All
subsequent clusters have been extracted with a random selection of seed’s URIs.
:x1 :x2 :x3 "female" i"Lnasuergenaanitne.l"e@tittere;
n
o
i
t
p
i
r
c
s
e
d
:
c
d
"Laurea in
giurisprudenza;
avvocato"@it
e
r
a
t
n
e
m
a
l
r
a
P
o
i
c
i
f
f
u
f
i
r
:
d
o |P |
c
2
81
44
:x4 148
398
7
10 Notice that, although the resources are the same as in the above example, we here
adopt a di↵erent criterion for selecting relevant triples.
This paper proposed a new and deductive strategy for clustering collections of
RDF resources on the basis of the informative content shared by their
descriptions expressed in form of generalized RDF triples. The clustering mechanism
relies on the computation of the CS [
The evaluated execution time of the whole clustering approach, together with the clustering results in terms of provided informative content, seem to support the e↵ort spent in designing and implementing the clustering strategy.
Part of the future work will be devoted to the extension of CS definition and computation to other entailment regimes, to the investigation on (as general as possible) criteria for the selection of relevant triples, aimed at the combined optimization of both description expressiveness and computational complexity, and to a comparative experimental evaluation involving the definition of a metric for clusters quality assessment. Acknoledgements We acknowledge support of projects “A Knowledge based Holistic Integrated Research Approach” (KHIRA - PON 02 00563 3446857) and “Enhance Risk Management through Extended Sensors” (ERMES - PON01 03113/F3).