Semantic Movie Recommendations - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 131

Table 5.1 The table shows the size of the entity and relationship sets of our movie dataset
Entity type

Number Relationship Number of edges Graph density

of entities movies ↔

Actors 95,321 Actors 231,742 1.80E-05

Directors 4,060 Directors 10,155 4.24E-06

Genres 20 Genres 20,809 9.80E-06

Tags 5,297 Tags 51,795 2.09E-05

Locations 187 Locations 49,167 2.30E-05

Countries 72 Countries 10,197 4.80E-06

Users 760 Users (train) 525,318 2.42E-04

Movies 65,133 Users (test) 330,280 1.52E-04

Discussion : In this section we explained our semantic datasets for the movie
recommendation domain. We presented the dataset we use for learning a semantic
movie recommender and discussed the properties of the dataset. The dataset com-
prises encyclopedic knowledge as well as rating knowledge describing the user's
preferences. Data collected from different sources are combined in one large graph
and stored in a uniform semantic data format. All nodes ( entities ) in the created
graph are identified by unique uniform resource identifiers (URIs). The dataset sta-
tistic (Table 5.1 ) shows, that the properties of the entity sets and the relationships
sets highly differ (according to number of elements and according to the density of
the relationships sets). Thus, the dataset allows us to analyze how our approach can
handle the heterogeneity in large semantic datasets.

5.3.4 Challenges and Requirements for Learning
Recommenders

Having created one large semantic knowledge graph that aggregates data from het-
erogeneous sources, we define an approach for learning a recommendation strategy.
The challenges of creating a powerful semantic recommender are: (1) The hetero-
geneity of the aggregated sub-graphs according to the number of nodes and edges,
(2) the sparsity of sub-graphs, and (3) the diversity of noise in the aggregated sub-
datasets. In addition, the sub-graphs may use different edge types and schemes for
assigning edge labels requiring a domain-specific model reflecting the heterogeneity
of the aggregated graphs. In the following paragraphs we discuss the challenges in
detail and explain approaches for the processing of heterogeneous semantic data.