Exploiting Linked Data-based Personalization Strategies for
Recommender Systems
Gabriela Oliveira Mota Da Silva
1,2 a
, Lara Sant’Anna do Nascimento
1 b
and Frederico Ara
´
ujo Dur
˜
ao
1 c
1
Institute of Computer Science, Federal University of Bahia, Avenida Milton Santos, s/n, Salvador, Brazil
2
Federal Institute of Education, Science and Technology of Bahia, Avenida Centen
´
ario, 500, Jacobina, Brazil
Keywords:
Recommender Systems, Linked Data, Personalization, Semantic Similarity, Feature Selection.
Abstract:
People seek assertive and reliable recommendations to support their daily decision-making tasks. To this end,
recommendation systems rely on personalized user models to suggest items to a user. Linked Data-driven
systems are a kind of Web Intelligent systems, which leverage the semantics of links between resources in the
RDF graph to provide metadata (properties) for the user modeling process. One problem with this approach
is the sparsity of the user-item matrix, caused by the many different properties of an item. However, feature
selection techniques have been applied to minimize the problem. In this paper, we perform a feature selection
preprocessing step based on the ontology summary data. Additionally, we combine a personalization strategy
that associates weights with relevant properties according to the user’s previous interactions with the system.
These strategies together aim to improve the performance and accuracy of the recommender system, since
only latent representations are processed by the recommendation engine. We perform several experiments on
two publicly available datasets in the film and music domains. We compare the advantages and disadvantages
of the proposed strategies with non-personalized and non-preprocessed approaches. The experiments show
significant increases in top-n recommendation tasks in Precision@K (K=5, 10), Map, and NDCG.
1 INTRODUCTION
Asking for recommendations is part of everyday life
for many people in different areas, whether it is shop-
ping, reading a book, listening to a song, or even look-
ing for a romantic couple. Thus, one searches for a
reliable source of recommendation so that the object
sought is following their tastes (Ricci et al., 2015). To
this end, recommender systems (RS) increasingly rely
on personalized user models by exploiting metadata
from the items that a user has previously interacted
with to suggest similar items they might be interested
in. The user model plays a significant role in many
personalized applications. For example, an RS ana-
lyzes the 1 to 5-star ratings given by a user to movies
and music in the past to establish their general inter-
ests in art. The system then evaluates and suggests
other items that it expects the user will genuinely like
in the future.
a
https://orcid.org/0000-0001-8799-5642
b
https://orcid.org/0000-0001-7428-9992
c
https://orcid.org/0000-0002-7766-6666
Linked Data (LD) or Linked Open Data (LOD)
is a powerful source of diverse data structured under
open Semantic Web standards, such as Resource De-
scription Framework (RDF) and SPARQL (Berners-
Lee, 2009). DBpedia
1
is a notable example of a LOD
set built by a strong community effort to extract struc-
tured information from Wikipedia and make this in-
formation openly available on the Web.
In the past years, LD datasets such as DBpedia
have been proposed as a valuable source of informa-
tion to increase the predictive power of RSs (Di Noia
et al., 2018). In such Linked Data-driven RS, links
between resources in the RDF graph play the primary
role of providing metadata (properties or features) for
the user modeling process and the recommendation
model (Passant, 2010; Di Noia et al., 2012; Piao and
Breslin, 2016). These systems use a semantic simi-
larity algorithm that calculates the degree of match-
ing between pairs of Linked Data resources. Because
RDF represents data as a graph, these algorithms, in
general, count the number of direct and indirect links,
1
https://www.dbpedia.org/
226
Da Silva, G., do Nascimento, L. and Durão, F.
Exploiting Linked Data-based Personalization Strategies for Recommender Systems.
DOI: 10.5220/0011591300003318
In Proceedings of the 18th International Conference on Web Information Systems and Technologies (WEBIST 2022), pages 226-237
ISBN: 978-989-758-613-2; ISSN: 2184-3252
Copyright
c
2022 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
the length of the path between two resources, or their
place in the hierarchy of classes (Passant, 2010; Piao
et al., 2016; Cheniki et al., 2016).
One problem of this approach is the sparsity of
the user-item matrix, provided by the numerous dif-
ferent properties one item could have. However, vari-
ous feature selection techniques have been applied to
discard irrelevant properties and minimize the prob-
lem. These techniques filter properties that can cause
the system’s reliability to decrease and only consider
the most relevant subset of the original data in the rec-
ommendation process step.
Nevertheless, some effort has been made to select
subsets of features that seem more helpful in com-
puting similarity between items of a graph dataset
(Catherine and Cohen, 2016; Musto et al., 2016),
reducing the matrix dimension. Recent studies cite
several methods to find and exclude or rank features
based on diverse goals — see Section 4 for examples.
This work investigates LD-based personalization
strategies to provide more representativeness to the
user model. The proposed algorithm associates
weights to the properties of items, according to their
relevance to the user. The relevance is calculated by
identifying among the items that the user has pre-
viously liked (user model), which properties link to
the same item repeatedly. We assume that properties
that appear more often in a particular user’s model
are more important for the calculation of future rec-
ommendations. We also perform a filtering prepro-
cessing step in which we apply an ontology-based
summarization method (Di Noia et al., 2018) to fil-
ter properties according to the domains. In summary,
these strategies together aim to reduce the dimension-
ality of the spaces and enhance the recommender sys-
tem performance.
We perform several rounds of experiments with
two different domain datasets - MovieLens and
LastFM, enriched with DBpedia semantic informa-
tion, to evaluate the system’s effectiveness and tune
its parameters. The similarities are calculated with
the Personalized Linked Data Semantic Distance
(PLDSD) method (da Silva et al., 2019). We present
the advantages and disadvantages of the proposed
strategies by comparison with non-personalized and
non-preprocessed approaches and discuss the results
in terms of performance and accuracy.
The remainder of the paper is organized as fol-
lows: Section 2 explains about Linked Data driven
RS; Section 3 discusses feature selection ad other per-
sonalization methods; Section 4 lists the work related
to this research; Section 5 introduces and describes
the proposed personalization approaches; Section 6
sets forth the experimental evaluation in the context
of two LD-based recommender systems simulation;
Section 7 includes the discussion on the results and
some points of improvement; and Section 8 concludes
the paper and provides suggestions for future work.
2 LD-BASED RECOMMENDER
SYSTEMS
Recommender Systems (RSs) are software tools and
techniques that aim to solve the information overload
problem by suggesting items that are most likely of
interest to a particular user (Ricci et al., 2015). Lit-
erature essentially discusses two recommender meth-
ods: Content-Based (CB), in which the similarity of
items is calculated by analyzing features associated
with the compared items; and Collaborative Filter-
ing (CF), in which recommendations to one user are
based on items that other users with similar tastes
liked in the past. Regardless of the chosen technique,
the user model plays a central role in recommender
systems since it encodes the user’s preferences and
needs (Ricci et al., 2015). For instance, in consider-
ing a CB approach, the user can be profiled directly
by its ratings of items. Thus, in a LD enriched Rec-
ommender System context, resources and their rela-
tions are mixed to the user model (ratings over items)
to calculate similarity and generate personalized user
recommendations (da Silva et al., 2019).
The data that describe items in an RS are derived
from many sources, most often private databases be-
longing to companies. However, the availability of
open sources of knowledge is growing, increasing the
emergence of semantic-based applications, for exam-
ple, Semantic-Aware SRs (de Gemmis et al., 2015).
Linked Open Data (LOD) is a widely known project
that aims to connect open datasets over the Web,
and facilitate their use by applications (Bizer et al.,
2009). The LOD project consists of 1,255 datasets
with 16,174 links, as of May 2020
2
, that covers an
extensive collection of statements related to resources
such as people, places, songs, movies and so on. A
typical LOD dataset is DBpedia, which incorporates
links to other datasets, such as DailyMed, Geonames,
and LinkedGeoData. As these additional links are
provided, applications can exploit knowledge from
interconnected datasets for developing semantic ap-
plications.
Semantics-aware applications rely on an RDF
graph architecture. The RDF statement (also
known as triple) has the form of subject-predicate-
object, each part uniquely identifiable by Uni-
2
https://lod-cloud.net/
Exploiting Linked Data-based Personalization Strategies for Recommender Systems
227
Figure 1: A DBpedia movie context example.
form Resource Identifiers (URIs). For example,
in the DBpedia movie segment shown in Figure
1, dbr:The Avengers is the subject of one triple,
dbo:director is the predicate, and dbr:Joss Whedon
is the object, forming the sentence: The Avengers
was directed by Joss Whedon. Another triple links
the subject dbr:Alien Ressurection to the same ob-
ject dbr:Joss Whedon by a different property
dbo:writer forming, in turn, the sentence: Alien
Ressurection was written by Joss Whedon.
This work focuses on CB Recommender Systems,
which compute the similarity between items in the
system by comparing their characteristics (also called
features). For instance, in a movie Recommender
System, algorithms reason about the degree of sim-
ilarity between two movies by comparing their direc-
tors, actors, the main subject, and so on. Although
several studies address different ways to build recom-
mendations, RS researchers still discuss some prob-
lems. One is the sparsity problem, when there is
much less feedback data in comparison to the whole
data matrix size; or the well-known cold-start prob-
lem, that happens when a new item or a new user is
added to the system. Another common issue is how
to offer more personalized recommendations to users.
LD are often used to add semantics to the system in
order to help address those well-known RS problems
(Di Noia et al., 2012; de Gemmis et al., 2015; Joseph
and Jiang, 2019).
In LD-based RSs, items’ features are expressed
by links between nodes in the RDF graph, that rep-
resent the resources’ properties. Thus, similarity al-
gorithms need to consider the semantics brought by
those links. Different similarity measures were pro-
posed to handle this task through LD-based systems
(Passant, 2010; Di Noia et al., 2012; Piao and Breslin,
2016). Most of them can manage the sparsity prob-
lem well. Nevertheless, one typical characteristic of
the previous work on LD similarity measures is that
most research considers all the links on an RDF graph
as having equal importance. Retaking the movie RS
context, genre is considered an important feature to
the movie domain. One user although likes to watch
only to recent released titles. Given this hypotheti-
cal scenario, is it correct that algorithms consider the
property genre as having the exact weight of the re-
lease year of a movie? Based on this issue, we will
discuss some ways to rank features regarding a given
RS context and/or select the best combination of prop-
erties that fits a particular user model.
3 FEATURE SELECTION AND
PERSONALIZATION
Many machine learning tasks, such as recommender
systems, demand a preprocessing step to filter a
smaller and more significant portion of the dataset.
When working with Linked Data, a single item in the
original dataset may have hundreds or thousands of
variables (as many as the number of links), causing
the feature matrix to be oversized. Studies covering
this issue focus mainly on two branches of research:
selecting subsets of features that are useful to build a
good predictor Feature Selection (FS), and rank-
ing all potentially relevant variables in a context
Feature Ranking (FR) (Guyon and Elisseeff, 2003).
One common way of performing FS tasks is
through hand-crafted work: manually selecting the
most relevant features according to simple heuristics
or domain knowledge. Nevertheless, there are at-
tempts to automatize the process, for example, Musto
et al. (Musto et al., 2016) assess the impact of some
classic feature selection techniques on recommenda-
tions accuracy, such as Principal Component Analysis
(PCA), Information Gain Ratio (GR), PageRank (PR)
and Support Vector Machines (SVM).
In an LD-Based Recommender System, proper-
ties are considered as features of a given node in the
knowledge graph. In a different approach, Di Noia et
al. (Di Noia et al., 2018) show how LD-based sum-
marization can drive FS and FR tasks by comparing
an automated feature selection method based on on-
tology data summaries with other classical ones, like
manual selection.
WEBIST 2022 - 18th International Conference on Web Information Systems and Technologies
228
Although FS and FR preprocessing techniques are
commonly applied in recommended models, most of
them consider links as having equal importance from
the user’s point of view. In this work, we further per-
form an FR task personalized. We not only use on-
tology summarization to select the most relevant fea-
tures from the LD graph, but we rank them by adding
weights to each link according to the user model.
This step is called graph personalization and is re-
sponsible for automatically ranking the properties that
most influenced the ratings given by a user of a rec-
ommender system. For example, in a movie RS, the
genre, artists, and director properties are usually con-
sidered important by automated FS algorithms such
as ontology summarization. However, if a particular
user has given positive ratings to movies of various
genres and directors, but most are short movies, our
personalization method will classify the movie length
as more important than the other properties, giving it
a higher weight.
4 RELATED WORK
This work addresses the problem of building a recom-
mender system over a LOD dataset, where resources’
properties form sparse matrices. Literature is plenty
of work that draws on a feature-based definition of
Linked Data, such as Meymandpour & Davis (Mey-
mandpour and Davis, 2016) which proposes a gener-
alized information content-based approach with sys-
tematic assessment of semantic similarity between
entities. They extensively reviewed existing seman-
tic similarity measures and proposed a hybrid method
made of feature-based and statistical techniques.
One example of a semantics-aware application
is when a Knowledge Graph (KG) is used to pro-
vide better user recommendations (Catherine and Co-
hen, 2016). In most works following this line of re-
search, the authors enrich benchmark databases with
semantics given from the interconnections between
KG nodes. Musto et al. (Musto et al., 2016) studies
the impact of using knowledge coming from Linked
Data on the overall performance of a recommenda-
tion algorithm. They reduced the matrices dimension
by testing several feature selection techniques against
two datasets. They investigate whether the integra-
tion of LD-based features improves the effectiveness
of the algorithm and to what extent the choice of dif-
ferent feature selection techniques influences its per-
formance in terms of accuracy and diversity.
Furthermore, Di Noia et al. (Di Noia et al., 2018)
show how LD-based summarization can drive fea-
ture selection tasks by proposing a fully automated
method based on semantics provided by the ontol-
ogy. They evaluate the accuracy of each strategy used
and compare them to classic methods such as man-
ual selection and statistical distribution methods. Fi-
nally, they aggregate diversity to the recommender
system by exploiting the top-k selected features. In
the inverse direction, Van Rossum & Frasincar (van
Rossum and Frasincar, 2019) investigate the incorpo-
ration of graph-based features into path-based simi-
larities. They proposed two normalization procedures
that adjust user-item path counts by the degree of cen-
trality of the nodes connecting them. It’s based on the
idea that a user liking one movie tells us more about
the popularity of this movie than about the particular
user. Our approach has similar premises, however, the
methodology diverges as we exploit the link informa-
tion from the perspective of each specific user.
Some studies were made in this line of user-based
personalization. However, approaches do not em-
brace graph-based systems. On Singh et al. (Singh
et al., 2018), user-based collaborative filtering is used
to generate recommendations by utilizing both items
and user preferences based on splitting criteria for
movie recommendation applications. Every single
item and every single user are split into two virtual
items and two virtual users based on contextual val-
ues. The recommender technique is then applied to
the new dataset of split items and users. A new ap-
proach is proposed by Yi et al. (Yi et al., 2018) by
combining the tasks of rating prediction from a rating-
based system with a review-based RS. In this manner,
they made a user-item rating relation from latent fea-
ture representations and fuse it to the user-item review
relation extracted from users’ reviews.
Latest research addresses new approaches to im-
prove the user experience in RS, as in Blanco. et al.
(Blanco. et al., 2021). They propose a recommender
recovery solution with an adaptive filter to deal with
the failed recommendations, which after a recommen-
dation failure, filters out all items that are similar to
the one disliked by the user. The objective is to keep
the user engagement and allow the recommender sys-
tem to become a long-term application. Same as Yi
et al., this approach is not adapted to LD-based sys-
tems. Another interesting approach is the use of long-
tail algorithms to generate recommendations that can
lead users to explore less popular but highly relevant
items in the RDF graph (de Sousa Silva et al., 2020).
Our approach aims to improve the user experience as
well, with the advantage of being LD-driven. Gan
et al. (Gan et al., 2021) proposed an EM-model that
alternates between a general item diversity learning
and knowledge graph embedding learning for user
and item representation, which helps to achieve better
Exploiting Linked Data-based Personalization Strategies for Recommender Systems
229
results in comparison to the state-of-art baselines on
datasets MovieLens and Anime. Although this work
takes advantage of auxiliary information along with
historical interactions between user and item from
knowledge graphs, it does not cover a personalization
method.
Cao et al. (Cao et al., 2019) proposed a research
based on the idea that a KG has commonly miss-
ing facts, relations, and entities. Thus, they argue
that it is crucial to consider the incomplete nature of
KG when incorporating it into recommender system.
The main difference between the related work cited
above and our approach fundamentally is that neither
of them really leverages the knowledge present in the
user model (KG) to make personalized recommen-
dations. Among the related works, the one that ap-
proaches most closely the idea of this work is Cao et
al., because they explore the KG trying to understand
the reasons why a user likes an item. They provide
an example that if a user has watched several movies
directed by (relation) the same person (entity), it is
possible to infer that the director relation plays a crit-
ical role when the user makes the decision, thus help
to understand the user’s preference at a finer granu-
larity (Cao et al., 2019). We also started our research
from this idea and developed a weighting algorithm
that ranks properties by importance from the user’s
point of view. Before the ranking step, we also ad-
dress the concept of KG completion, although we use
the rationale of direct and indirect links coming from
Passant (Passant, 2010). We generate the missing di-
rect links by counting the indirect links between the
items the user interacted with, as we demonstrate be-
low.
5 PROPOSAL
This section is dedicated to explaining the strategies
proposed in this work, beginning with the user graph
personalization, and going through how the weights
are applied together with a similarity measure for rec-
ommending items to the user.
5.1 Notation
A user model or user profile is how a recommender
system represents the users’ preferences about items,
as discussed in Section 2. In this work, we model
the user profile as part of the Linked Data graph that
describes the system’s domain. Therefore, we as-
sume a set of RDF resources identified by their URI
R = {r
1
,r
2
,...,r
n
}, and a set of users identified by
their URI U = {u
1
,u
2
,...,u
m
}, in which each instance
of u
k
is also itself a resource in the LD graph, in other
words, U R. Figure 2 shows an example of a user
resource u
1
= dbr:User1, which is represented in yel-
low, to distinguish it from other resources.
We also assume a set of properties P =
{p
1
, p
2
,..., p
h
} that link together pairs of resources
from R molding triples in the form of T =
{t
1
,t
2
,...,t
q
}. Thus, as an RDF triple is a state-
ment in the format hsubject, predicate, objecti, then
t
i
= hr
a
, p
j
,r
b
i T means that there is an instance
of a property p
j
P linking the subject r
a
R to
the object r
b
R, like in hdbr:Toy Story, dbo:director,
dbr:John Lasseteri, back to Figure 2.
Following the DBpedia notation, we elaborated
one property p
x
= dbr : liked that represents all the
positive feedback from users over the items, with
p
x
P. For example, a triple t
1
= hu
1
,dbr : liked, r
1
i
means that the user u
1
U has rated the item r
1
R
as a positive feedback.
In essence, a LOD graph is a directed graph
G = (R,P,T ), that encompass the sets of resources
R = {r
1
,r
2
,...,r
n
}, properties P = {p
1
, p
2
,..., p
h
} and
triples T = {t
1
,t
2
,...,t
q
}, in combination with a set
of special resources called users U = {u
1
,u
2
,...,u
m
},
and one special property p
x
= dbr : liked, which to-
gether provide the notation for modeling the user’s
preferences over the dataset items.
5.2 Graph Personalization
Our proposed strategy is composed of a user model
personalization approach (da Silva et al., 2019),
and of an ontology-based feature selection approach
adapted for LD contexts from Di Noia et al. (Di Noia
et al., 2018). In the personalization step, any resource
r
a
from R that is linked to a user u
k
by the particular
property p
x
is considered as part of the user model.
These resources are always of the same type, which
is the main class of the domain dataset. For exam-
ple, Figure 2 models a portion of a movie dataset,
where we can notice the resources from the movie
class highlighted in lavender color. The three of them
are in the model because the user u
1
= dbr:User1 has
liked them through the property p
x
= dbr : liked.
The personalization method intends to assign
weights to every property p
j
from P in the graph G
that links a user model item colored in lavender in
Figure 2 and represented in equations by r
a
— to any
other resource mostly the ones colored in green,
represented by r
b
. The goal is to rank properties by
applying weights so that the aspects of the film that
influenced the user the most are at the top of the list. It
means that we assign higher weights to the properties
p
j
that occur more among the liked movies. Equation
WEBIST 2022 - 18th International Conference on Web Information Systems and Technologies
230
1 shows how the weights calculation is made for each
user u
k
.
W (p
j
,u
k
) =
1
1 +
t
i
Freq(r
b
)
q
(1)
Where t
i
= hr
a
, p
j
,r
b
i T , and Freq(r
b
) calculates
the frequency of the resource r
b
, by iterating over the
triples t
i
= hr
a
, p
j
,r
b
i T with i varying from 1 to q.
Since the resource r
a
always refers to an item that
was positively evaluated by a user u
k
, the W equation
is performed for every p
j
until a rank of properties is
generated from the perspective of a particular user u
k
.
In future recommendations, the ranking will be taken
into account when calculating the similarity of new
items. In resume, the graph personalization process
is based on the concept that the more one property is
associated with appreciated items, the greater the im-
portance of this property in future recommendations
for the user.
5.3 Personalized LDSD
In this work, we show that ranking the most rele-
vant features from the user’s point of view, instead
of considering just the context model, increases the
effectiveness of the RS. The classic LDSD (Passant,
2010) measure considers every link p
j
as having the
same weight on calculations. In order to make the
recommender system personalized, we add the func-
tion W (p
j
u
k
) to the PLDSD Equation. As W repre-
sents the weight to one property for a given user’s
point of view, its value is multiplied with each func-
tion C(p
j
,r
a
,r
b
), which are iterated to the j number
of properties linking r
a
and r
b
, as shown in Equation
2.
The PLDSD distance counts direct and indirect
links from resource r
a
to resource r
b
and vice versa.
All functions C(p
j
,r
a
,r
b
) compute whether there is a
link p
j
between resources r
a
and r
b
. They return 1
in case there is, 0 otherwise. Whenever there is an n,
the function is calculating the total number of links
between a resource r
a
or r
b
to all other resources. C
d
only takes into consideration direct links from r
a
to r
b
;
reverse links, from r
b
to r
a
, count as different links.
The function C
ii
represents indirect incoming links
and returns 1 only if there is a resource r
c
that satisfies
both hp
j
,r
a
,r
c
i and hp
j
,r
b
,r
c
i, 0 if not. In the case of
returning 1, a virtual direct link is created between the
two resources that share the incoming link. The func-
tion C
io
represents indirect outgoing links and equals
to 1 only if there is a resource r
c
that satisfies both
hp
j
,r
c
,r
a
i and hp
j
,r
c
,r
b
i, 0 if not. In the case of re-
turning 1, a virtual direct link is created between the
two resources that share the outgoing link.
In Figure 2 the highlighted property dct:subject
was created because The Amazing Spider Man and
The Avengers share the same outgoing link. Simi-
larly, the property dbo:product was created because
The Avengers and Toy Story share the same ingoing
link. It is possible to infer from the figure that these
two properties will have a greater weight after the W
equation is calculated, because they appear more fre-
quently in the graph, influencing the PLDSD result.
5.4 Semantic Feature Selection
We adapted the ABSTAT Feature Selection method
(Di Noia et al., 2018) by applying the LDSD mea-
sure to it and implemented a preprocessing step that
filters out the most important features for each do-
main. That method originally uses the Jaccard index,
a well-known but non-specific similarity measure for
LD-based systems.
ABSTAT is the name of the original implementa-
tion (Di Noia et al., 2018), which consists of an au-
tomatic FS method, the ontology-based data summa-
rization framework. This framework identifies pat-
terns and frequencies and computes cardinality de-
scriptors from the RDF triples of one domain. The au-
thors conduct experiments in three different domains:
movies, music and books.
6 EXPERIMENTAL EVALUATION
The goal of the experimental evaluation is to validate
the hypothesis that personalized user models can lead
to more effective recommendations in Linked Data-
based systems. To prove this, we implement a person-
alization strategy that weighs and ranks each property
in the user model according to the user’s past interac-
tions with the system, called PLDSD.
We performed two comparative experiments
against the classical LDSD method, one using a
movie RS and one using a music RS. Results show
greater accuracy of the system when using PLDSD
instead of the non personalized LDSD. Additionally,
we run the PLDSD method with and without the pre-
processing step the semantic Feature Selection
method as additional validation experiments. We
demonstrate that using a filter technique appropriate
to the application domain — both by being LD-based
sim
PLDSD
(r
a
,r
b
,u
k
) =
1
1 +
j
C
d
(p
j
,r
a
,r
b
).W (p
j
u
k
)
1+logC
d
(p
j
,r
a
,n)
+
j
C
d
(p
j
,r
b
,r
a
).W (p
j
u
k
)
1+logC
d
(p
j
,r
b
,n)
+
j
C
ii
(p
j
,r
a
,r
b
).W (p
j
u
k
)
1+logC
ii
(p
j
,r
a
,n)
+
j
C
io
(p
j
,r
a
,r
b
).W (p
j
u
k
)
1+logC
io
(p
j
,r
a
,n)
(2)
Exploiting Linked Data-based Personalization Strategies for Recommender Systems
231
Figure 2: Example of weight calculation in PLDSD.
and by taking into account the semantics of the do-
main elements also increases the accuracy of the
recommendations when combined with the personal-
ization strategy. We critically discuss all these results
in Section 7.
The experiments were implemented in Java using
JENA, an open source framework for building Se-
mantic Web and Linked Data applications
3
. We also
store similarity values and other data in a MySQL
database and leverage the linked data environment
with the Virtuoso Server
4
. The complete implemen-
tation project has been made available on an Anony-
mous Github under the url https://anonymous.4open.
science/r/lodweb-pldsd-40BD/. Researchers inter-
ested in replicating and conducting further compar-
ative studies can find the necessary documentation at
this location.
6.1 Dataset Setup
Two datasets were configured to run the experimen-
tal evaluations. The first dataset is the Movielens 1M
dataset
5
, which has: 1 million ratings from 6000 users
on 4000 movies (1-5 stars rating). The Dataset rating
density is 4.26%. From this start point, we modeled a
database that stores the user model users, movies,
and user ratings over movies, and also stores the con-
tent of movies in the form of RDF triples. For this
purpose, we used the MappingMovielens2DBpedia
project
6
, which provides RDF identifiers URIs
3
https://jena.apache.org/
4
https://virtuoso.openlinksw.com/
5
https://grouplens.org/datasets/movielens/1m/
6
https://github.com/sisinflab/LODrecsys-datasets/blob/
master/Movielens1M/MappingMovielens2DBpedia-1.2.tsv
for each movie in MovieLens 1M. Therefore, this ini-
tial setup allows us to access resources and all their
connections from DBpedia through SPARQL queries
online.
The second dataset is the Last.FM Million Song
7
.
It contains listening information from almost 2 thou-
sand users about approximately 1 million songs.
Nonetheless, we take the listening data summarized
to user-artist for this work, rather than considering
the songs played. For this reason, during the dataset
setup step, we used another set of data released dur-
ing the HETRec 2011 Workshop and published as an-
other project of raw data mapping to Linked Data
8
.
As is in the movie mapping project, this one provides
DBpedia URIs for each musical artist or band in the
Last.FM dataset.
6.2 Methodology
In this work, we simulate two recommender sys-
tems that retrieve the Top-N recommendations to the
user: a Movie RS and a Music RS. User prefer-
ences about movies in the MovieLens dataset are ex-
pressed on a 5-star Likert scale. Some work on the
MovieLens dataset use methods for making the user
ratings binary (Di Noia et al., 2012; Musto et al.,
2016). The PLDSD methodology considers 3 differ-
ent scenarios to represent a user’s positive feedback:
PLDSD
rat=5
, that only ratings equal to 5 are consid-
ered; PLDSD
rat>=4
that considers movies rated as 5
and 4; and PLDSD
rat>=3
that takes ratings equal to 5,
7
http://millionsongdataset.com/lastfm/
8
https://github.com/sisinflab/LinkedDatasets/blob/
master/last fm/mappingLinkedData.tsv
WEBIST 2022 - 18th International Conference on Web Information Systems and Technologies
232
4 and 3 as positive feedback. Previous work (da Silva
et al., 2019) demonstrated that the PLDSD
rat=5
strat-
egy performed the most accurate results for all sce-
narios. For this reason, we adapted our methodology
to consider ratings 5 as positive feedback and all the
other ratings — 1, 2, 3, and 4 — as negative feedback.
Similarly, in the Last.FM dataset, user preferences
about artists are expressed through the number of
times that each user has listened to their songs. Con-
sequently, the concept of positive feedback for this
work had to be constructed by isolating an expres-
sive number that represents a user listening to a well-
liked artist throughout their daily routine. We chose
the number 500 to represent positive feedback and
therefore artists tracks heard less than 500 times are
considered negative feedback given by a user. Based
on this, we have built the user models from the set
of items positively evaluated by each user. We also
personalized the user models by applying the PLDSD
methodology as explained in section 5.
The personalization step results in lists of ranked
properties according to the user’s past choices in the
system. These lists may contain all the properties in-
volved in the user model, or they may be limited by a
cutoff point, which is the value of the parameter k in
the ontology summarization approach (section 3). In
other words, the parameter k in the feature selection
preprocessing step defines how many properties will
be ranked by the personalization step.
From this point on, we will refer to movies and
musical artists simply as items, since the methodol-
ogy applies to any domain and we would like to de-
scribe it as generally as possible. First, we take items
that the user has given positive feedback and build the
training dataset or the dataset of known data. Then we
take a subset of 90 items randomly (discarding those
that are already in the user model subset) and build
the test dataset. After that, we perform the PLDSD
algorithm on the test dataset and also perform the
LDSD algorithm as the baseline method. The user
model is then used twice during each round of exper-
iments: first to perform the weigh preprocessing step
described in Subsection 5; and secondly as the vali-
dation dataset to which we compare the ranked result
list coming from the recommendation method.
Moreover, we performed some statistical tests to
assess the significance of the experimental results.
First, we ran the chi-squared goodness-of-fit test over
the ranked lists of movies and music. As this test re-
vealed the normal distribution of the data, we opted
to apply a paired T-test using a p-value << 0.0001 to
each PLDSD sample with 3 variations of k values
— against its correspondent baseline method (LDSD)
results’ sample — with 3 more variations of k values.
As the length of the sample varies according to how
many items the user has rated, we ran statistical tests
for each user tested.
6.3 Metrics
We compute the results under 3 evaluation metrics to
assess the quality of the generated ranks. These met-
rics assess the relevance and accuracy of the similarity
algorithm. In other words, they measure how accu-
rately a system can predict similar items to a specific
user.
Precision@K - Precision measures the fraction of
retrieved items from a dataset that are relevant, as
shown in Equation 3. Precision@K is a variation
that represents the amount of relevant results at
the position K (Equation 4). We have adopted in
our evaluation two classic values of K: 5 and 10
(Davoodi et al., 2013).
Precision =
|RelevantItems RetrievedItems|
|RetrievedItems|
(3)
Precision@K =
|RelevantItems@K|
K
(4)
Mean Average Precision (MAP) - It is a com-
bination of averages that has good discrimination
and stability (Manning et al., 2008). MAP is an
extension of the Average Precision (AP) which
takes the average of all AP as shown in Equation
5.
MAP(Q) =
1
|Q|
+
|Q|
j=1
1
m
j
m
j
k=1
Precision(R
jk
) (5)
Given the set {d
1
,...,d
m
j
} of relevant documents
for an information need q
j
Q and R
j
k is the set
of ranked results from the top result until docu-
ment d
k
.
Normalized Discounted Cumulative Gain
(NDCG) - It is a metric based on the notion that
items in a rank have varying degrees of relevance,
either once relevant items appear in a low-rank
position or when less relevant items appear in a
greater position. (J
¨
arvelin and Kek
¨
al
¨
ainen, 2002).
Thereby, the gained value (without discounts) is
obtained as the relevance score of each item is
progressively summed from the rank position 1
to n. Discounted Cumulative Gain (DCG) is then
obtained through Equation 6, which penalizes
highly relevant documents appearing lower on a
search result list as the graded relevance value
is logarithmically reduced proportionally to the
result’s position.
Exploiting Linked Data-based Personalization Strategies for Recommender Systems
233
DCG[i] =
(
CG[i], if i < b
DCG[i 1] + G[i]/log
b
i, if i b.
(6)
DCG values can be compared to the theoretically
best possible score vector, the ideal vector, which is
represented by Equation 7 for relevance scores 0 and
1, where m is the number of relevant items.
BV [i] =
(
1, if i m
0, otherwise.
(7)
Normalizing the values is a required step when
comparing two techniques under DCG. This is made
by dividing the DCG vectors by the corresponding
ideal DCG vectors (iDCG). The NDCG is then calcu-
lated using Equation 8 which, likewise Precision@K,
only evaluates the top k results (Manning et al., 2008).
NDCG(k) =
DCG(k)
iDCG(k)
(8)
7 RESULTS
This section presents and discusses the results of each
evaluation metric, considering the methodology and
objectives proposed by this work. Before the tests, we
built the user models from the set of items positively
evaluated by each user. The concept of positive rating
was constructed differently for each domain dataset,
movies, and music, according to the methodology ex-
plained in section 6.
The personalization step results in lists of ranked
properties, that are used to feed the recommender sys-
tem, which runs two similarity algorithms: LDSD and
Personalized LDSD (PLDSD). The ranked lists may
be limited by a k value due to the feature selection pre-
processing step. As we consider 3 k values for each
strategy — in addition to the turn without limiting the
value of k —, and 2 similarity strategies, the experi-
mental evaluation consists of 8 rounds of testing for
each user on the movie dataset and another 8 rounds
for each user on the music dataset.
7.1 Discussion
In order to facilitate the understanding of the results
tables, we use the acronym LDSD to represent the
recommender model built under plain LDSD, and
the acronym PLDSD to represent the recommender
model that considers our personalized approach, as
presented in Section 5. Tests that consider the feature
selection step are identified by the value assigned to k
in the result tables: Table 1 and Table 2. For exam-
ple, PLDSD k=10 means that results encompass the
personalization method and the preprocessing step,
that selects the 10 most relevant properties. On the
other hand, LDSD k=0 demonstrates the results for
when neither customization nor feature summary are
applied.
Table 1 summarizes the average (Avg) results of
a hundred users from the movie dataset, considering
the LDSD scenarios with and without feature selec-
tion and PLDSD with and without feature selection.
In the movie scenario, each of the 100 user models is
composed of at least 7 and at most 52 positively rated
movies. The test set is composed of 200 not evaluated
movies. The value of k varies from 10 to 535, which is
the maximum number of properties for the movie do-
main in DBpedia. LDSD and PLDSD without feature
selection (k=0) are used for comparison purposes.
Results of Table 1 show a statistically significant
increase on all metrics when applying the PLDSD
with k=10. We can notice that the LDSD approach
presents worse results in comparison with the person-
alized and filtered rounds.
Table 1: Results for the movie dataset considering diverse
scenarios.
Strategy Precision@5 Precision@10 MAP NDCG
LDSD k=0 0.450 0.450 0.501 0.743
LDSD k=10 0.500 0.450 0.565 0.798
LDSD k=100 0.450 0.425 0.550 0.751
LDSD k=535 0.450 0.425 0.499 0.744
PLDSD k=0 0.450 0.450 0.511 0.757
PLDSD k=10 0.550 0.450 0.582 0.809
PLDSD k=100 0.550 0.450 0.581 0.794
PLDSD k=535 0.550 0.425 0.504 0.783
Table 2 summarizes the average (Avg) results of
a hundred users from the music dataset, considering
LDSD scenarios with and without feature selection
and PLDSD with and without feature selection. In the
music scenario, each of the 100 user models consists
of at least 14 and at most 50 positively rated artists.
The test set is composed of other 200 not evaluated
artists. The value of k varies from 10 to 512, which is
the maximum number of properties for the music do-
main in DBpedia. LDSD and PLDSD without feature
selection (k=0) are used for comparison purposes.
As with the movie domain, the music results
shown in Table 2 also achieved higher values when
applying PLDSD with k=10. And again the results
with LDSD without personalization and without fil-
ters show worse results among the rounds of experi-
ments.
The results for the movie dataset experiments
show that MAP and NDCG metrics obtained the best
values. Additionally, those values are higher when the
similarities are calculated using the feature selection
step. Figure 3 presents the MAP and NDCG values
for both LDSD and PLDSD approach with different
WEBIST 2022 - 18th International Conference on Web Information Systems and Technologies
234
Table 2: Results for the music dataset considering diverse
scenarios.
Strategy Precision@5 Precision@10 MAP NDCG
LDSD k=0 0.433 0.450 0.491 0.622
LDSD k=10 0.462 0.450 0.520 0.671
LDSD k=100 0.425 0.425 0.503 0.621
LDSD k=512 0.420 0.425 0.488 0.603
PLDSD k=0 0.436 0.450 0.509 0.635
PLDSD k=10 0.470 0.556 0.535 0.731
PLDSD k=100 0.451 0.445 0.508 0.647
PLDSD k=512 0.450 0.440 0.500 0.620
values of k. This graphical representation emphasizes
the significant growth of approximately 7% in the
NDCG for LDSD values and of 8% in the NDCG for
PLDSD values. Concerning the MAP metric, exper-
iments achieved 13% of growth in the LDSD rounds
and 14% in the NDCG for PLDSD values when ap-
plying the feature selection step.
The results of Precision@5 and Precision@10 in
the movie experiments show positive and negative os-
cillations between values, highlighting a positive gain
of Precision@5 when calculating PLDSD with the
feature selection. Furthermore, all metrics results are
higher when the number of properties is reduced in
the pre-selection step, both for LDSD and PLDSD,
especially when k = 10.
Figure 3: NDCG and MAP results versus the k value of
pre-selected features from the movie domain.
Figure 4: NDCG and MAP results versus the k value of
pre-selected features from the music domain.
The results of the experiments with the music
dataset also showed improvements, although not as
expressive as the movie domain, especially for the
MAP metric. Figure 4 shows a growth of 8% in the
NDCG for LDSD values and of 15% in the NDCG for
PLDSD values. Regarding the MAP metric, exper-
iments achieved 6% of growth in the LDSD rounds
and 5% in the NDCG for PLDSD values when apply-
ing the feature selection step.
Originally, the ABSTAT method behave differ-
ent depending on the selected knowledge domain
(Di Noia et al., 2018). A comparative analysis leads
us to conclude that ABSTAT summaries are strongly
grounded in the ontological nature of the knowledge
graph, while our approach emphasizes user prefer-
ences drawn from their previous interactions with the
system. This means that PLDSD applied to different
domains will perform similarly, although it may vary
slightly. Nevertheless, the combined use of the two
techniques has been shown to be more efficient than
either one separately.
However, further analysis is required to investi-
gate if the nature of the music subject does not allow
an accurate prediction of its properties weigths. The
maximun number of properties retrieved from DB-
pedia is similar to both domains 535 and 512
although the modeling is very different. A musical
artist can be linked to many different songs by the
property is dbo:artist of, while an actor is apt to be
linked to fewer movies by the property is dbo:star of.
For example, the band The Rolling Stones is linked
to over 300 songs, while Anthony Hopkins, the actor
with one of the most solid careers, could act in only
137 movies so far.
Conversely, the overall performance of the system
shows a significant reduction in the processing time,
both for the movie and song datasets, since the num-
ber of properties being computed is reduced by ap-
proximately 80% when considering K = 10. Taking
the example of user #1 who positively rated 246
artists — the computation time was reduced from 240
to 67 minutes by adding the selection step, a reduction
of 72%.
7.2 Points of Improvement
During the evaluation phase, one difficulty was to de-
fine the number of properties that should be used in
the pre-selection process, that is, to define the value of
k. Due to this fact, the tests were performed consider-
ing three distinct values for k. We defined the values
so that the algorithm could explore a low (k = 10),
a medium (k = 100) and a high (k = max) value of
k for pre-selection. The highest value of k is the
maximum number of properties returned by the pre-
selection step with the dataset used in the tests, which
is k = 535 for movies and k = 512 for music.
Although most of the results were more relevant
when k = 10, both for LDSD and PLDSD, it is neces-
sary to establish a method that identifies the optimal
amount of properties to be analyzed for each data set.
Exploiting Linked Data-based Personalization Strategies for Recommender Systems
235
One suggestion for future work is to apply the Elbow
Curve, a popular method for finding the optimal num-
ber of clusters when working with the K-Means clas-
sification algorithm (Kaufman and Rousseeuw, 1990).
We propose to investigate the possibility of finding the
ideal value to K by adapting the Elbow Method, with
the purpose of improving the results personalization
strategies of this work.
8 CONCLUSION
In this paper, we proposed personalization strategies
for LD-based recommender systems. We use a user
modeling process that analyses the past interactions
of the user with the system to rank the properties that
are used in the recommender model. After ranking
the properties, we applied the Personalized Linked
Data Semantic Distance (PLDSD) similarity measure,
which generates a rank of items to recommend to the
user. We run experiments comparing the PLDSD re-
sults to the classic LDSD measure using 3 different
metrics. We also performed comparative experiments
using an adapted implementation of an LD-oriented
feature selection strategy, so that only the most rele-
vant properties for the system were considered in part
of the calculations.
The evaluation results show the best values for
PLDSD combined with a k = 10 choice of feature
selection strategy, that outperforms the unweighted
and not filtered baseline method LDSD. We can state
that this work achieved the goal of obtaining better
accuracy and performance of LD-based RS when us-
ing movie and music datasets from DBpedia. An im-
provement in the results was noticed when the num-
ber of items evaluated was increased and the number
of selected properties was reduced with the applica-
tion of the filtering step. The results of the PLDSD
metrics combined with the filter properties stood out
from the others in all the tests performed.
As future work, we aim to test our model us-
ing other LD-based similarity measures in order to
compare and determine which one performs better.
We also plan to conduct more studies regarding the
feature selection task, by using other LD-driven ap-
proaches and comparing them to the baseline meth-
ods used so far. Another possible future work is to
evaluate this approach using a cross-domain dataset,
which would enable the development of multi-domain
recommendations for general use in linked datasets.
ACKNOWLEDGEMENTS
This work was partially funded by the Coordenac¸
˜
ao
de Aperfeic¸oamento de Pessoal de N
´
ıvel Superior -
Brasil (CAPES) – Grant number 001.
REFERENCES
Berners-Lee, T. (2009). Linked-data design issues. w3c de-
sign issue document. The World-Wide Web Consor-
tium W3C.
Bizer, C., Heath, T., and Berners-Lee, T. (2009). Linked
data - the story so far. Int. J. Semantic Web Inf. Syst.,
5(3):1–22.
Blanco., J., Ge., M., and Pitner., T. (2021). Recom-
mendation recovery with adaptive filter for recom-
mender systems. In Proceedings of the 17th Inter-
national Conference on Web Information Systems and
Technologies - WEBIST,, pages 283–290. INSTICC,
SciTePress.
Cao, Y., Wang, X., He, X., Hu, Z., and Chua, T.-S. (2019).
Unifying knowledge graph learning and recommen-
dation: Towards a better understanding of user pref-
erences. In The World Wide Web Conference, WWW
’19, page 151–161, New York, NY, USA. Association
for Computing Machinery.
Catherine, R. and Cohen, W. (2016). Personalized recom-
mendations using knowledge graphs: A probabilis-
tic logic programming approach. In Proceedings of
the 10th ACM Conference on Recommender Systems,
RecSys ’16, pages 325–332, New York, NY, USA.
ACM.
Cheniki, N., Belkhir, A., Sam, Y., and Messai, N. (2016).
Lods: A linked open data based similarity measure.
2016 IEEE 25th International Conference on En-
abling Technologies: Infrastructure for Collaborative
Enterprises (WETICE), pages 229–234.
da Silva, G. O. M., Dur
˜
ao, F. A., and Capretz, M. (2019).
Pldsd: Personalized linked data semantic distance for
lod-based recommender systems. In Proceedings of
the 21st International Conference on Information In-
tegration and Web-Based Applications and Services,
iiWAS2019, page 294–303, New York, NY, USA. As-
sociation for Computing Machinery.
Davoodi, E., Kianmehr, K., and Afsharchi, M. (2013). A se-
mantic social network-based expert recommender sys-
tem. Applied Intelligence, 39(1):1–13.
de Gemmis, M., Lops, P., Musto, C., Narducci, F., and Se-
meraro, G. (2015). Semantics-aware content-based
recommender systems. In Ricci, F., Rokach, L., and
Shapira, B., editors, Recommender Systems Hand-
book, pages 119–159. Springer US, Boston, MA.
de Sousa Silva, D. V., de Oliveira, A. C., Almeida, F., and
Dur
˜
ao, F. A. (2020). Exploiting graph similarities with
clustering to improve long tail itens recommendations.
In Proceedings of the Brazilian Symposium on Multi-
media and the Web, WebMedia ’20, page 193–200,
New York, NY, USA. Association for Computing Ma-
chinery.
WEBIST 2022 - 18th International Conference on Web Information Systems and Technologies
236
Di Noia, T., Magarelli, C., Maurino, A., Palmonari, M.,
and Rula, A. (2018). Using ontology-based data sum-
marization to develop semantics-aware recommender
systems. In Gangemi, A., Navigli, R., Vidal, M.-E.,
Hitzler, P., Troncy, R., Hollink, L., Tordai, A., and
Alam, M., editors, The Semantic Web, pages 128–144,
Cham. Springer International Publishing.
Di Noia, T., Mirizzi, R., Ostuni, V. C., Romito, D., and
Zanker, M. (2012). Linked open data to support
content-based recommender systems. In Proceedings
of the 8th International Conference on Semantic Sys-
tems, I-SEMANTICS ’12, pages 1–8, New York, NY,
USA. ACM.
Gan, L., Nurbakova, D., Laporte, L., and Calabretto, S.
(2021). Emdkg: Improving accuracy-diversity trade-
off in recommendation with em-based model and
knowledge graph embedding. In IEEE/WIC/ACM In-
ternational Conference on Web Intelligence and In-
telligent Agent Technology, WI-IAT ’21, page 17–24,
New York, NY, USA. Association for Computing Ma-
chinery.
Guyon, I. and Elisseeff, A. (2003). An introduction to
variable and feature selection. J. Mach. Learn. Res.,
3:1157–1182.
J
¨
arvelin, K. and Kek
¨
al
¨
ainen, J. (2002). Cumulated gain-
based evaluation of ir techniques. ACM Trans. Inf.
Syst., 20(4):422–446.
Joseph, K. and Jiang, H. (2019). Content based news rec-
ommendation via shortest entity distance over knowl-
edge graphs. In Companion Proceedings of The 2019
World Wide Web Conference, WWW ’19, pages 690–
699, New York, NY, USA. ACM.
Kaufman, L. and Rousseeuw, P. J. (1990). Finding Groups
in Data: An Introduction to Cluster Analysis. John
Wiley.
Manning, C. D., Raghavan, P., and Sch
¨
utze, H. (2008).
Evaluation in information retrieval, page 139–161.
Cambridge University Press.
Meymandpour, R. and Davis, J. G. (2016). A semantic sim-
ilarity measure for linked data. Know.-Based Syst.,
109(C):276–293.
Musto, C., Lops, P., Basile, P., de Gemmis, M., and Semer-
aro, G. (2016). Semantics-aware graph-based recom-
mender systems exploiting linked open data. In Pro-
ceedings of the 2016 Conference on User Modeling
Adaptation and Personalization, UMAP ’16, pages
229–237, New York, NY, USA. ACM.
Passant, A. (2010). Measuring semantic distance on linking
data and using it for resources recommendations. In
AAAI Spring Symposium: Linked Data Meets Artifi-
cial Intelligence. AAAI.
Piao, G., Ara, S. s., and Breslin, J. G. (2016). Comput-
ing the semantic similarity of resources in dbpedia for
recommendation purposes. In Qi, G., Kozaki, K., Pan,
J. Z., and Yu, S., editors, Semantic Technology, pages
185–200, Cham. Springer International Publishing.
Piao, G. and Breslin, J. G. (2016). Measuring semantic dis-
tance for linked open data-enabled recommender sys-
tems. In Proceedings of the 31st Annual ACM Sym-
posium on Applied Computing, SAC ’16, pages 315–
320, New York, NY, USA. ACM.
Ricci, F., Rokach, L., and Shapira, B. (2015). Recom-
mender systems: introduction and challenges. In Rec-
ommender systems handbook, pages 1–34. Springer
International Publishing.
Singh, M., Sahu, H., and Sharma, N. (2018). A Person-
alized Context-Aware Recommender System Based on
User-Item Preferences, pages 357–374. Springer In-
ternational Publishing, Singapore.
van Rossum, B. and Frasincar, F. (2019). Augmenting
lod-based recommender systems using graph central-
ity measures. In Bakaev, M., Frasincar, F., and Ko,
I.-Y., editors, Web Engineering, pages 19–31, Cham.
Springer International Publishing.
Yi, J., Huang, J., and Qin, J. (2018). Rating prediction in
review-based recommendations via adversarial auto-
encoder. In 2018 IEEE/WIC/ACM International Con-
ference on Web Intelligence (WI). IEEE.
Exploiting Linked Data-based Personalization Strategies for Recommender Systems
237