Semantic Movie Recommendations - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 132

5.3.4.1 The Graph-Based Knowledge Representation

We represent knowledge in a large graph consisting of nodes and edges. For ensuring
that all knowledge of the graph can be taken into account and for simplifying the
processing, we make sure that the graph is connected. Disconnected components
might be connected with the main component by adding edges with a very lowweight
(e.g., in a music genre graph each node may be connected with a general genre node).
Entities with the same semanticmeaning retrieved fromdifferent sources should have
the same URI. If the considered sources use different URIs for semantically identical
entities, we unify the URIs using “mapping” edges, such as owl:sameAs .

5.3.4.2 Mapping Edge Labels to Similarity Scores

Recommender systems usually compute recommendations based on the estimated
relatedness between users and items [ 9 , 16 ]. For ensuring that knowledge from
different sources can be combined, we model the relatedness between entities with
numerical similarity values. Thus, wemap the labels of edges that indicate a similarity
(such as liked or user has bought an item ) to numerical values (e.g., on a scale

).
The resulting distributions of similarity scores must be analyzed when computing the
recommendations. That is why we discuss scaling and weighting models adapting
the similarity scores for the needs of the applied recommender algorithms in the next
section.

[ 0 , 1 ]

5.3.4.3 Scaling Models

We define the relatedness of two entities in a graph based on the edges connecting
these entities. Initially, we assign for each semantic edge connecting the nodes n i and
n j (having an influence on the relatedness of two nodes) a similarity score w ij =

1;
if two nodes n i and n j are not connected by a semantic edge, we assign the weight
w ij = 0. Thus, we get an adjacency matrix containing only the values 0 and 1. Due

to the fact that most graphs are sparse, we suggest using sparse matrixes for storing
the graphs, keeping only the nonzero weighted edges.
Since different semantic edge types may have a different impact on the relatedness
of two nodes, we define the scaling factor for each semantic edge type. We compute
the adapted similarity score by multiplying the initial score with a scaling factor. The
scaling factor is defined based on expert knowledge. For example, the semantic edge
“user u has bought item i ” usually implies a higher relatedness than “user u has read
the description of item i ”.
The node degree (the number of nodes directly connectedwith the respective node)
is another import aspect that should be considered when computing the relatedness
score of two nodes. Entities, highly connected, often represent popular entities ( liked
by almost everyone). Edges connecting popular item nodes with user nodes often
do not contain much information about individual user preferences. Thus, dependent