Semantic Movie Recommendations - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 138

5.4.3 Ensemble-Based Recommender Approaches
Instead of creating one universal recommender for one large graph, we consider learn-
ing several different recommenders for sub-graphs and combining them in an ensem-
ble. Ensembles have the advantage that the recommenders in the ensemble have a
lower complexity and recommenders can be incrementally updated. In addition, new
algorithms can be integrated in order to cover new aspects. For example, results from
recommenders optimized for encyclopedic knowledge and recommenders trained
on personalized ratings can be combined in an ensemble. The strength of ensemble
approaches is that different algorithms can be combined considering the heterogene-
ity within the graph.
The disadvantage of ensembles consists in the overhead for managing different
algorithms in the ensemble and in the additional effort for combining the suggestions
from different algorithms.
In summary, ensemble approaches allow us the flexible combination of optimized
recommender algorithms. Ensemble approaches often enable improving the recom-
mendation quality as well as the trust in the system [ 23 ].
5.5 A Semantic Movie Recommender
We evaluate the developed approach in a web-based movie-recommender applica-
tion. The system has been created based on our semantic movie dataset aggregating
data from MovieLens, Freebase, and IMDb. The recommender system suggests
users interesting movies based on user-defined lists of favorite movies. The recom-
mendations are computed using agent ensembles combining the suggestions from
different semantic graphs.
In this section we explain the system architecture, present the graphical user
interface, and discuss the advantages for the users.
5.5.1 The System Architecture
We implement the movie recommendation system as an open, extensible web appli-
cation. The user interface is implemented using Grails 6 [ 25 ] running on an Apache
Tomcat 7 web server.
The system architecture is visualized in Fig. 5.4 . Each semantic relationship
set is wrapped by one agent allowing us updating and adding semantic relation-
ship. The recommender agents are optimized according to the specific proper-
ties of the wrapped semantic relationship sets. In order to provide personalized
6 https://grails.org/ .
7 http://tomcat.apache.org/ .