Information Technology Reference
In-Depth Information
7.5.4 Evaluation of the 'New User and New
Application' Scenario
The application is created by randomly selecting 5,000 users from our crawled data.
These users represent users who already use the application. The test users, which are
different from the 5,000 users, are also chosen from the user dataset containing only
users with an interest in jazz or swing music. This was done to augment the cold start
problem as most users in our dataset have a “Pop” taste. The initial user profiles are
enriched using data from Freebase. Figure
7.17
presents the results for the different
algorithms described in Sect.
7.5.3
. The results show that the enrichment has a huge
positive impact on the recommendation quality. Both approaches using the enriched
profiles (CF using only enriched profiles and
CF with standard profiles
combined
with
CF with enriched profiles
) clearly outperform the
standard CF
and
CF
Most
Popular
for user profiles of size 1-4. For users with a user profile size of five,
the
CF with enriched profiles
is slightly inferior than the
standard CF
.
Most Popular
and
Random
recommendation have no impact at all. Using only the
Most Popular
recommendations does not work as the selected test users were only interested in
swing and jazz music as the common taste in the randomly selected dataset is on pop
music. Thus, the list of
Most Popular
recommendations consists of pop artist and
does not contain any swing or jazz artists.
Figure
7.18
shows the change in recommendation quality on a percentage basis
compared to the
CF with standard profiles
. The usage of our enrichment approach
+
Fig. 7.17
Cluster-based prediction: Explanation of cluster-based enrichments using automatically
generated genre cluster
Search WWH ::
Custom Search