Information Technology Reference
In-Depth Information
Fig. 7.18 Cluster-based prediction: Explanation of cluster-based enrichments using automatically
generated genre cluster
improves the recommendation quality by over 90 % for very small profiles (sizes 1
and 2) and over 40 % for the bigger profiles (size 4 and 5).
7.5.5 Evaluation of Facebook and LastFM
The evaluation for the second scenario, the 'New user and large application' scenario,
is done separately for Facebook and LastFM to compare if there are differences in
a Social Network and a distinguished music recommendation service like LastFM.
The evaluation covers three different user subsets:
1. Recommendations based on the complete dataset.
2. Recommendations for users who have an uncommon taste. This is similar to the
swing and jazz user sets used in evaluation in Sect. 7.5.4 .
3. Recommendations for users who mostly like popular artists.
The split between users with an unusual taste and users with a common popular
taste is done based on the average deviation of popular artists in a user's profile.
The popularity of an artist is computed based on the distribution in the Facebook and
LastFM datasets. The initial user profiles (with one-five items) are enriched with five
additional items from Freebase, so that the user profiles given to the collaborative
filtering recommender have a size of six-ten items.
Figures 7.19 and 7.20 show the evaluation results on both datasets. Results on
the Facebook dataset show that CF with enriched user profiles does not improve the
recommendation quality compared to the standard CF . The enrichment even leads
to a reduced precision.
This is expected as our enrichment approach adds mostly 'popular' entities from
Freebase to the user profile, meaning that the enrichment can blur the user profile
and make the user profile less personal. As explained, the enrichment algorithm
takes the degree of a node in the Freebase graph into account. Thus, mostly popular
Search WWH ::




Custom Search