Authors:
Gabriela Oliveira Mota Da Silva
1
;
2
;
Lara Sant'Anna do Nascimento
1
and
Frederico Araújo Durão
1
Affiliations:
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 Centenário, 500, Jacobina, Brazil
Keyword(s):
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.
(More)