Semantic Movie Recommendations - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 135

dataset complexity (resulting in a reduced sparsity) and for extracting the most rel-
evant information (resulting in a reduced level of noise). Clustering approaches as
well as dimensionality reduction methods can be applied to detect irrelevant data
allowing us to reduce the complexity and the sparsity of the dataset [ 5 ].
Discussion : In this sectionwe discussed how to create and optimize a unified graph
based on data retrieved from several different sources. We presented approaches for
the domain-specific scaling of edge weights and models for aggregating the edge
weights of a complex path. Based on the optimized semantic graphs, recommender
models can be computed. In the next section we analyze algorithms for computing
recommendations considering the computational complexity, the recommendation
accuracy, and the ability for providing explanations.

5.4 Semantic Recommender Approaches

Having defined a unified graph with numerical edge weights, we analyze methods
for computing recommendations. In the following sections we discuss the differ-
ent recommender approaches and analyze the respective strengths and weaknesses.
In our analysis we focus on ( 1 ) Memory-based recommenders, ( 2 ) Model-based

recommenders, and ( 3 ) ensemble-based approaches.

5.4.1 Memory-Based Recommender

Approaches for graph-based recommenders can be classified according to the
data structures internally used. Memory-based recommenders compute suggestions
directly on the graph. This simplifies the adding and removing of data due to the
fact that there is no internal model that must be adapted to new data. Memory-based
recommenders for semantic graphs compute suggestions directly on the graph using
graph-search algorithms, such as Branch and Bound [ 24 ]. Since the run-time
complexity of these algorithms grows exponentially with the considered path length,
memory-based approaches usually consider only entities reachable by relatively short
paths. It is assumed that the relevant entities can be found in the near environment
around the input entities.
Memory-based approaches have the advantage that updates in the semantic graph
immediately affect the computed recommendations. Consequently, no additional
resources for model updates are needed. Moreover, memory-based recommenders
can provide human readable explanations, visualizing the nodes and edges considered
while computing the recommendations. In most scenarios the path length is limited
so that the explanations are not too complex ensuring that the explanations are under-
standable for the users. A visualization of an explanation generated by a memory-
based music recommender is shown in Fig. 5.3 . The example explanation visualizes
how starting from an input node (e.g., from the user profile), a recommendation is
computed. Starting from the movie node King Kong (2005) , the recommender