RDF Data Clustering based on Resource and Predicate Embeddings
Siham Eddamiri
1
, El Moukhtar Zemmouri
1
and Asmaa Benghabrit
2
1
LM2I Laboratory, Moulay Ismail University, ENSAM, Meknes, Morocco
2
LMAID Laboratory, Mohammed V University, ENSMR, Rabat, Morocco
Keywords:
Machine Learning, Linked Data, RDF, Clustering, Word2vec, Doc2vec, K-means.
Abstract:
With the increasing amount of Linked Data on the Web in the past decade, there is a growing desire for
machine learning community to bring this type of data into the fold. However, while Linked Data and Machine
Learning have seen an explosive growth in popularity, relatively little attention has been paid in the literature
to the possible union of both Linked Data and Machine Learning. The best way to collaborate these two fields
is to focus on RDF data. After a thorough overview of Machine learning pipeline on RDF data, the paper
presents an unsupervised feature extraction technique named Walks and two language modeling approaches,
namely Word2vec and Doc2vec. In order to adapt the RDF graph to the clustering mechanism, we first applied
the Walks technique on several sequences of entities by combining it with the Word2Vec approach. However,
the application of the Doc2vec approach to a set of walks gives better results on two different datasets.
1 INTRODUCTION
In recent years, the Web evolved from a web of do-
cuments to a web of data (Bizer et al., 2009). In fact,
the Web emerges from a global information space of
interlinked documents to one where both documents
and data are linked. Allowing to provide a solid foun-
dation for this evolution to be best practices for pu-
blishing and interlinking structured data on the Web
known as Linked Data principles (Bizer, 2011).
Meanwhile, the combination of Linked Data and
the field of Machine Learning hasnt paid much, while
both fields have seen an exponential growth in popu-
larity in the past decade. Therefore, the best solution
is to focus on the Resource Description Framework
(RDF) which is in the lowest layers in the Semantic
Web stack. Thus, to understand RDF, the ML resear-
cher do not need to know about ontologies, reasoning
and SPARQL to have the benefit.
As a matter of fact, the RDF data-model (Bloem
et al., 2014), is not suited for traditional machine lear-
ning algorithms. In fact, traditional machine learning
techniques requires a tabular as input and not on large
interconnected graphs such as RDF graphs. There-
fore, expressing such data in RDF will add many rela-
tions and concepts on top of its native structure, which
will add much inferencing, harmonization, and acces-
sibility, and also a new impediments and challenges.
Moreover, in the semantic web, a dataset doesnt
separate any longer by instances, or emerge from a
single learning task, even the standard methods of
evaluation dont adapt well. Therefore, a new pipeline
for RDF must be adopted in order to provide a com-
prehensive framework and to fit the common steps for
traditional machine learning techniques. While the
transformation to RDF is reversible, reconstructing
the original data requires manual effort or domain-
specific methods in order to suit the pipeline. Hence,
the preprocessing of RDF data to a similar form is
required to process large amounts of RDF data by ge-
neric methods.
In this paper, we introduce an overview for ma-
chine learning on RDF data, and a set of common
techniques for data pre-processing, in order to answer
most of common tasks in machine learning. In this
regard, we first convert the RDF graph into a set of
sequences using two approaches for generating graph
walks. Then, we adopt two language modelling ap-
proaches for latent representation of entities in RDF
graphs to train the sequence of entities, finally we use
K-means to group those entities vectors.
The rest of this paper is structured as: in section
2, we give some necessary preliminaries followed by
related works in Section 3. Section 4 describes our
approaches and then we present in section 5 an evalu-
ation of our propositions. Finally, we conclude with a
summary.
Eddamiri, S., Zemmouri, E. and Benghabrit, A.
RDF Data Clustering based on Resource and Predicate Embeddings.
DOI: 10.5220/0007228903670373
In Proceedings of the 10th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2018) - Volume 1: KDIR, pages 367-373
ISBN: 978-989-758-330-8
Copyright © 2018 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
367

2 PRELIMINARIES
The Resource Description Framework (RDF) is a lan-
guage for representing metadata about Web resources
(Manola and Miller, 2004). It has been introduced
and recommended by the World Wide Web Consor-
tium (W3C) as a fundamental building block of Lin-
ked Data and the Semantic Web. The central idea of
RDF is to enable the encoding, exchange and reuse of
structured metadata.
The atomic construct of RDF data model are state-
ments about resources in the form (subject, predicate,
object) called triples. A triple (s, p, o) in a set of tri-
ples T is composed by a subject s that is a resource
identified by URI, a property/predicate p also identi-
fied by URI and an object o that is the value of the
property. An object can be either another resource or
a literal. An RDF resource can also be a blank node
(Manola and Miller, 2004).
This simple model of assertions enables RDF to
represent a set of triples as a directed edge-labelled
graph G = (V, E) where s and o are nodes in V , and
(s, o) ∈ E is an edge oriented from node s to node o
and labeled with predicate p.
In this paper, we adopt the RDF formalization
introduced by (Tran et al., 2009) given as follow:
Definition 1 (RDF Data-Graph). An RDF data-
graph G representing a set of triples T is defined as a
tuple G = (V, E, L) where:
• V = V
E
∪V
l
∪V
b
define a finite set of vertices as
the union of V
E
the set of resource vertices, V
l
the
set of literal vertices and V
b
the set of blank node
vertices.
V = {v | ∃ x, y (v, x, y) ∈ T ∨ (x, y, v) ∈ T }
• E is a finite set of directed edge e(v
1
, v
2
) that con-
nect the subject v
1
and the object v
2
:
E = {(v
1
, v
2
)|∃x(v
1
, x, v
2
) ∈ T ∨ (v
2
, x, v
1
) ∈ T }
• L = L
V
∪ L
E
is a finite set of labels as the union
of L
V
a set of vertex labels and L
E
a set of edge
labels:
– If v ∈ V
l
then L(v) is a literal value.
– If v ∈ V
E
then L(v) is the resource correspon-
ding URI.
– If v ∈ V
b
then L(v) is set to NULL.
– If e ∈ E then L(e) is the property corresponding
URI.
L = l(v) = v|v ∈ V ∪ l(e) = e
This formalization fails when predicate terms are also
used in the subject or object position. Therefore, we
must distinct between the position of the predicate
s
1
s
2
s
3
s
4
p
1
p
2
p
3
p
4
p
5
(a)
s
1
s
2
s
3
s
4
p
1
p
2
p
3
p
4
p
5
(b)
Figure 1: (a) Example of RDF graph that is edge-labeled
multigraph; (b) The obtained RDF bipartite graph.
and the position of the subject and object, even if they
have the same label.
As a matter of fact, the main challenge of RDF
data which defined as an example of labeled mul-
tigraph (edge-labeled multigraph) is that there are
many edges between a pair of vertices. Therefore,
we transform an edge-labeled multigraph RDF to an
RDF bipartite graph (Figure 1) that incorporate state-
ments and properties as nodes into the graph, in order
to deliver a richer sense of connectivity than the stan-
dard directed labeled graph representations. To do so,
we opt to reify the edges by adopting the formaliza-
tion introduced by (de Vries et al.,2013), describing
an RDF bipartite graph as:
So, there is a vertex for each unique subject and
object. And for each triple we create two edges,
one for connecting subject to triple and other for
connecting triple to object. This definition has the
additional advantage that occur in applying any algo-
rithms that use labeled simple graphs, regardless of
whether or not they are constructed from a triple store.
Definition 2 (RDF Bipartite Graph). An RDF bipar-
tite data graph G is defined as a tuple (V, E, l) where:
• V = V
E
∪ V
l
∪ V
b
∪ V
T
a finite set of vertices as
the union of V
E
the set of resource vertices, V
l
the
set of literal vertices, V
b
the set of blank nodes
vertices, and V
T
the set of triple vertices.
V = {v | ∃ x, y (v, x, y) ∈ T ∨ (x, y, v) ∈ T } ∪ T
• E is a finite set of directed edges that connect a
subject v
1
to a triple and a triple to object v
2
:
KDIR 2018 - 10th International Conference on Knowledge Discovery and Information Retrieval
368

E = {(v
1
, v
2
)|∃x(v
1
, x, v
2
) ∈ T ∨ (v
2
, x, v
1
) ∈ T }
• L = L
V
∪ L
E
is a finite set of labels as the union
of L
V
a set of vertex labels and L
E
a set of edge
labels:
– If v ∈ V
l
then L(v) is a literal value.
– If v ∈ V
E
then L(v) is the resource correspon-
ding URI.
– If v ∈ V
b
then L(v) is set to NULL.
– If e ∈ E then L(e) is the property corresponding
URI.
L = l(v) = v|v ∈ V ∪ l(e) = e
So, there is a vertex for each unique subject and
object. And for each triple we create two edges, one
for connecting subject to triple and other for con-
necting triple to object. This definition has the additi-
onal advantage that occur in applying any algorithms
that use labeled simple graphs, regardless of whether
or not they are constructed from a triple store.
3 RELATED WORK
The combination of Machine Learning and Sematic
Web focuses on making an RDF graph as input to an
existing ML algorithm, by creating a generic pipeline
which will help to have some standardized methods.
We have drawn an example pipeline (Figure 2), to
summon up from RDF to an ML model by the fol-
lowing steps:
Pre-processing consists on extracting the most ef-
ficient and more relevant information from RDF data,
by dealing with three common problems: (1) Clea-
ning data based on detecting and correcting or even
removing a vertex/property/triples from an RDF data;
(2) Hub removal based on the principle that node
which have many links to other nodes add a little in-
formation about an instance, hence we could relabel
the hub vertex by the combination of edges and vertex
labels (Bloem et al., 2014),on one hand, or removing
all the frequent pairs from the graph using a parameter
k and replacing the pair that have the lowest frequency
by the concatenation of the labels (de Vries, 2013) on
the other hand; (3) Relabeling based on giving a new
label to a vertex (blank nodes or a hub vertex).
Instance Extraction consists on extracting sets of
subgraphs that are relevant to a given resources (ver-
tices) from an RDF graph. For this, there are several
approaches which we can categorize into three appro-
aches: (1) Immediate Properties/Nave Extraction is
a direct approach that consider the immediate pro-
perties of resources (Colucci et al., 2014; Ristoski
and Paulheim, 2016);(2) Concise Bounded Descrip-
tion (CBD) is the improvement over the immediate
properties which consider the types of nodes into ac-
count in the graph;(3) Depth Limited Crawling (DLC)
is limited the sub-graph by depth and simply traverse
the graph of certain number of steps from the starting
node, which could be traversing forward or/and bac-
kwards edges from the root node in the graph such
as (L
¨
osch et al., 2012; de Vries, 2013; Bach, 2008;
Zhao and Karypis, 2002).
Feature Extraction consists on transforming
RDF graph to features which reduce irrelevant and re-
dundant variables, while describing the data with suf-
ficient accuracy. The current state of the art for this
step given by Weisfeiler-Lehman (WL) graph kernel
(de Vries and de Rooij, 2015; Bach, 2008) aimed at
improving the computation time of the kernel while
applied RDF, which can be seen as a graph or as a le-
arner. The WL algorithm was adapted to compute the
kernel directly on the underlying graph, while main-
taining a subgraph perspective for each instance as
(L
¨
osch et al., 2012; de Vries, 2013).
Learning consists on feeding the feature vec-
tors or the subgraphs to a learner which we can di-
vide to two types : (1) Supervised Learning based
on training a data sample from labeled data source
(L
¨
osch et al., 2012); (2) Unsupervised learning is a
Self-Organizing neural network learn which based on
identifying a hidden pattern in unlabelled input data
(Grimnes et al., 2008; Colucci et al., 2014; Ristoski
and Paulheim, 2016).
4 THE METHODOLOGY
4.1 Overview
As a matter of fact, in case of RDF graphs, we con-
sider entities and relations between them instead of
word sequences. Therefore, we convert the RDF
graph data into sequences of entities (vertices + ed-
ges) in order to apply such approaches. After that, we
train these sequences to the neural language models
to generate vectors of numerical values in a latent fe-
ature space, then we cluster them by using K-means
algorithm in order to obtain groups. To do so, in this
paper we use two approaches for extracting instances
and two approaches for generating embedding vectors
from an RDF graphs.
4.2 Feature Vector using Word2vec
In this approach depicted in figure 3, for a given RDF
graph G = (V, E, l) we extract first the vertices (in-
stances) that we would cluster, then for each vertex
v we generate all walks of depth d starting with v by
RDF Data Clustering based on Resource and Predicate Embeddings
369

Figure 2: An overview of a Machine Learning pipeline for RDF.
the BFS (Breath-First Search) algorithm. The set of
sequences for graph G is the union of all the walks ge-
nerated by exploring the direct outgoing edges of the
root vertex v
r
, while for each explored edge r
1i
we ex-
tract the neighbourhood vertex, so for the depth d = 1
the sequences has the following structure v
r
, r
1i
, v
1i
and vice versa for the neighborhood vertex until the
depth d.
Definition 3: A Walks is defined as a sequence of
vertices and edges with a specific depth d as the fol-
lowing formalization :
Walks = {(v
i−1
, e
i
, v
i
) ∈ T |0 ≤ i ≤ d}
RDF Graph
Walks
Word2vec
Kmeans
Figure 3: The overall process to map the walks toward
word2vec.
For the training, we use word2vec (Mikolov et al.,
2013), to construct a vocabulary from the training file
and then learns high dimensional vector representa-
tions of words. according to the parameter in Table
1.
4.3 Feature Vector using Doc2vec
In this approach (Fig. 4), we use the same technique
as the first approach to generate the walks, rooted with
a specific vertex v
r
by a depth d. however, the diffe-
rence here is to gather all the walks that start with
the same vertex into a file, hence we will have for
each vertex a file that contain all the walks, a set of
walks, as defined at the beginning of 4.2, with a depth
d rooted with this vertex. Then we use doc2vec (Le
and Mikolov, 2014), which is inspired by word2vec
Table 1: Training parameters using Word2vec.
Parameters Explanations Default value
Train Name of input file Train.txt
Output Name of output file Vectors.bin
SG Choice of training model 0
0: Skip-gram model
1: CBOW model
Size Dimension of vectors 200/300
Negative Number of negative samples 10
Windows Distance between the 4
current and predicted word
for larger blocks of texts (documents) to learn embed-
dings and then learns high dimensional vector repre-
sentations of documents according to the parameter in
Table 2.
Definition 4 : A set of Walks is defined as the union
of walks with a specific vertex v
r
by a depth d as the
following formalization :
∑
Walks(v
r
) = ∪{(v
r
, e
r+1
, v
r+1
) ∈ T |0 ≤ i ≤ d}
RDF Graph
∑
Walks(v)
Doc2vec
Kmeans
Figure 4: The overall process to map the walks toward
word2vec.
5 RESULTS AND DISCUSSION
Evaluate both approaches on two datasets and two dif-
ferent feature extraction strategies (Walk and the set
of walks), combined with two different learning algo-
rithms (Word2vec and Doc2vec).
KDIR 2018 - 10th International Conference on Knowledge Discovery and Information Retrieval
370

Table 2: Training parameters using Doc2vec.
Parameters Explanations Default value
Train Name of input file Train.txt
Output Name of output file Vectors.bin
DM Choice of training model 0
0: distributed bag of words model
1: distributed memory model
Size Dimension of vectors 200/300
Negative Number of negative samples 10
Windows Distance between the 4
current and predicted word
5.1 Dataset
In order to evaluate our approaches, we use two ty-
pes of RDF Graphs, derived from existing RDF da-
tasets, as showed in Table 3. The value of a certain
property is used as a classification target. However,
first we specifiedd the number of clusters generated
for both datasets based on the preexisting classifica-
tion and calculated purity, Fmeasure.
The AIFB dataset represented in (Bloehdorn and
Sure, 2007) describes the AIFB research institute in
terms of its staff, research groups, and publications.
In total this dataset contains 178 persons that belong
to 1 of 5 research groups. However, one of these
groups has only 4 embers, which we ignore from the
dataset, leaving 4 groups. The goal of the experi-
ment is to training and grouping the 174 persons into
4 groups
The BGS dataset represented in (de Vries, 2013)
was created by the British Geological (BGS) Survey
and describes geological measurements in Great Bri-
tain in the form of Named Rock Units. For these na-
med rock units, we have two largest classes for litho-
genesis property, each class have 93 and 53 instances.
The aim of our experiment is to training and grouping
these 163 instances into 2 groups.
Table 3: The parameters of training command using
Word2vec.
Dataset Instances Clusters Source
AIFB 176 4 AIFB
BGS 146 2 BGS
5.2 Performance Measures
The effectiveness of the clustering step can be eva-
luated by two measures. They allow evaluating the
reliability of the clustering mechanism by comparing
the obtained clusters to known classes. The first mea-
sure is the F-measure, it is based on the fundamental
informational retrieval parameters, precision an recall
which are defined as :
P(i, j) =
n
i j
n
j
(1)
R(i, j) =
n
i j
n
i
(2)
Where n
i
is the number of the vertices of the class i,
n
j
is the number of the vertices of the cluster j and
n
i j
is the number of the class i in the cluster j. The
F-measure of a cluster and a class i is given by :
F(i, j) =
2P(i, j) ∗ R(i, j)
P(i, j) + R(i, j)
(3)
For the entire clustering result, the F-measure is com-
puted as following:
F =
∑
i
n
i
n
max
j
(F(i, j)) (4)
Where n is the total number of vertices. A better
clustering result is measured by the largest F-measure
(Steinbach et al., 2000). The second measure is the
purity of a cluster represents the ratio of the dominant
class in the cluster to the size of the cluster. Thus, the
purity of the cluster is defined as:
purity( j) =
1
n
j
max
i
(n
i j
) (5)
The global value for the purity is the weighted average
of all purity values. It is given by the following for-
mula:
purity =
∑
j
n
j
n
purity( j) (6)
A better clustering result is measured by the largest
purity value (Zhao et al.,2004).
5.3 Results
The results for the task of clustering on the two RDF
datasets are given in Tables 4 and 5. From the re-
sults we can observe that the combination of set of
walks with doc2vec approach outperforms all the ot-
her approach. More precisely, using the skip-gram
feature vectors of size 300 provides the best results
on the two datasets. The combination of the walks
with word2vec approach on all two datasets performs
closely to the standard graph substructure feature ge-
neration strategies, but it does not outperform them.
Doc2vec approach generate vectors description of an
instance based on all the walks related to this instance
while, word2vec approach consider each walk sepa-
rately. Hence, we observe that the skip-gram perform
RDF Data Clustering based on Resource and Predicate Embeddings
371

Table 4: The results of similar features clustering based on word2vec and doc2vec in terms of Purity.
AIFB Dataset BGS Dataset
Word2Vec Doc2Vec Word2Vec Doc2Vec
SG CBOW PV-DBOW PV-DM SG CBOW PV-DBOW PV-DM
Size = 200
d = 2 46.33 46.32 55.71 63.03 67.86 66.67 85.62 78.77
d = 3 49.15 47.46 62.50 73.58 60.36 57.14 76.71 72.60
d = 4 64.15 48.59 55.71 68.09 76.19 82.53 69.86 71.23
Size = 300
d = 2 47.46 48.02 79.67 72.77 60.00 66.67 86.99 78.08
d = 3 79.10 49.15 63.38 68.06 61.60 63.91 67.12 71.92
d = 4 47.46 48.02 64.79 73.50 78.12 61.90 77.40 71.92
Table 5: The results of similar features clustering based on word2vec and doc2vec in terms of F-measure.
AIFB Dataset BGS Dataset
Word2Vec Doc2Vec Word2Vec Doc2Vec
SG CBOW PV-DBOW PV-DM SG CBOW PV-DBOW PV-DM
Size = 200
d = 2 44.51 45.01 50.04 57.01 62.72 62.21 62.92 59.23
d = 3 44.21 45.72 58.40 73.70 60.08 63.55 71.11 59.28
d = 4 58.81 46.51 49.61 64.11 76.19 82.36 59.45 68.77
Size = 300
d = 2 44.52 45.84 56.51 64.73 60.00 59.48 63.49 58.86
d = 3 56.91 47.01 59.50 67.84 61.19 59.48 56.85 66.90
d = 4 44.13 46.51 59.80 63.70 69.10 57.45 71.90 69.54
better than the CBOW in the approach word2vec with
size 300 and the PV-DBOW perform better than PV-
DM in Doc2vec in term of purity.
6 CONCLUSIONS
In this paper, we have presented an overview of a Ma-
chine Learning pipeline on RDF data, and a set of
common techniques for data preprocessing, in order
to get solve the most common tasks in machine lear-
ning from linked data by requiring steps such as : Pre-
processing, instance extraction, feature extraction and
learning. Then we have identified two mean challen-
ges for performing clustering of Semantic Web data,
which are instances extraction instance from a large
RDF graph. Moreover, we have evaluated two dif-
ferent approaches on AFIB and BGS datasets. Af-
ter comparing these approaches, we opted for the set
of walks approach to generate an instance extraction
from RDF graph. Finally, we believe that we have
shown that clustering of RDF resources is an inte-
resting area, where many questions remain unanswe-
red. Our middle-term goal is to choose the optimal in-
stance extraction methods, distance metrics and clus-
tering algorithms to perform our RDF clustering.
REFERENCES
Bach, F. R. (2008). Graph kernels between point clouds.
In Proceedings of the 25th International Conference
on Machine Learning, ICML ’08, pages 25–32, New
York, NY, USA. ACM.
Bizer, C. (2011). Evolving the web into a global data space.
In Proceedings of the 28th British National Confe-
rence on Advances in Databases, BNCOD’11, pages
1–1, Berlin, Heidelberg. Springer-Verlag.
Bizer, C., Heath, T., and Berners-Lee, T. (2009). Linked
data - the story so far. International journal on se-
mantic web and information systems, 5(3):1–22.
Bloehdorn, S. and Sure, Y. (2007). Kernel methods for mi-
ning instance data in ontologies. In Aberer, K., Choi,
KDIR 2018 - 10th International Conference on Knowledge Discovery and Information Retrieval
372

K.-S., Noy, N. F., Allemang, D., Lee, K.-I., Nixon,
L. J. B., Golbeck, J., Mika, P., Maynard, D., Mizogu-
chi, R., Schreiber, G., and Cudr-Mauroux, P., editors,
The Semantic Web, volume 4825 of Lecture Notes in
Computer Science, pages 58–71. Springer.
Bloem, P., Wibisono, A., and de Vries, G. (2014). Simpli-
fying RDF data for graph-based machine learning. In
KNOW@LOD, volume 1243 of CEUR Workshop Pro-
ceedings. CEUR-WS.org.
Colucci, S., Giannini, S., Donini, F. M., and Sciascio, E. D.
(2014). A deductive approach to the identification
and description of clusters in linked open data. In
Proceedings of the Twenty-first European Conference
on Artificial Intelligence, ECAI’14, pages 987–988,
Amsterdam, The Netherlands, The Netherlands. IOS
Press.
de Vries, G. K. (2013). A fast approximation of the
weisfeiler-lehman graph kernel for rdf data. In Pro-
ceedings of the European Conference on Machine Le-
arning and Knowledge Discovery in Databases - Vo-
lume 8188, ECML PKDD 2013, pages 606–621, New
York, NY, USA. Springer-Verlag New York, Inc.
de Vries, G. K. D. and de Rooij, S. (2015). Substructure
counting graph kernels for machine learning from rdf
data. Web Semant., 35(P2):71–84.
Grimnes, G. A., Edwards, P., and Preece, A. (2008). In-
stance based clustering of semantic web resources.
In Bechhofer, S., Hauswirth, M., Hoffmann, J., and
Koubarakis, M., editors, The Semantic Web: Research
and Applications, pages 303–317, Berlin, Heidelberg.
Springer Berlin Heidelberg.
Le, Q. and Mikolov, T. (2014). Distributed representations
of sentences and documents. In Proceedings of the
31st International Conference on International Con-
ference on Machine Learning - Volume 32, ICML’14,
pages II–1188–II–1196. JMLR.org.
L
¨
osch, U., Bloehdorn, S., and Rettinger, A. (2012). Graph
kernels for rdf data. In Simperl, E., Cimiano, P., Pol-
leres, A., Corcho, O., and Presutti, V., editors, The Se-
mantic Web: Research and Applications, pages 134–
148, Berlin, Heidelberg. Springer Berlin Heidelberg.
Manola, F. and Miller, E. (2004). RDF primer. World Wide
Web Consortium, Recommendation REC-rdf-primer-
20040210.
Mikolov, T., Yih, S. W.-t., and Zweig, G. (2013). Linguis-
tic regularities in continuous space word representa-
tions. In Proceedings of the 2013 Conference of the
North American Chapter of the Association for Com-
putational Linguistics: Human Language Technolo-
gies (NAACL-HLT-2013). Association for Computati-
onal Linguistics.
Ristoski, P. and Paulheim, H. (2016). Rdf2vec: Rdf graph
embeddings for data mining. In Groth, P., Simperl, E.,
Gray, A., Sabou, M., Kr
¨
otzsch, M., Lecue, F., Fl
¨
ock,
F., and Gil, Y., editors, The Semantic Web – ISWC
2016, pages 498–514, Cham. Springer International
Publishing.
Steinbach, M., Karypis, G., and Kumar, V. (2000). A com-
parison of document clustering techniques. In In KDD
Workshop on Text Mining.
Tran, T., Wang, H., and Haase, P. (2009). Hermes: Data web
search on a pay-as-you-go integration infrastructure.
Web Semant., 7(3):189–203.
Zhao, Y. and Karypis, G. (2002). Evaluation of hierarchical
clustering algorithms for document datasets. In Pro-
ceedings of the Eleventh International Conference on
Information and Knowledge Management, CIKM ’02,
pages 515–524, New York, NY, USA. ACM.
RDF Data Clustering based on Resource and Predicate Embeddings
373