Collaborative Use of Features in a Distributed System for the Organization of Music Collections - Intelligent Music Information Systems: Tools and Methodologies - page 171

Figure 14. Different taxonomies created with Nemoz. The upper one structures the audio files according
to mood, the other according to the singer's voice.

(a) Mood taxonomy (b) Voice taxonomy

criterion. The underlying feature set contained
20 “allround” features (Mörchen, Ultsch, Thies,
Löhken, Nöcker, & Stamm, 2004) and 123 special-
ized features created by the approach described in
Mierswa and Morik (2005). We selected 10 base
features which performed well on a wide range
of learning tasks in order to allow for an efficient
comparison. We then generated base feature
weights for each learning task in the training set
and test set using a SVM (Rüping, 2000). The
average accuracy was measured for all learning
tasks in the test set. We used the base feature
set only, the selected set of optimal features for
each learning task, and features recommended by
case-based feature construction with a varying
number of predictors. We also measured the ef-
fort of finding a good subset of features in terms
of run-time and the number of times the inner
evaluation operator for feature selection approach
was invoked.
The result is shown in Table 5. As expected,
using the base features only, leads to the lowest
accuracy. Choosing an optimal set of features for
each task separately leads to the highest accuracy.
Using the features of the k most similar taxonomies

with respect to the similarity measure described
in this section performs in between. However,
this feature transfer efficiently achieves accuracy
close to the optimal one. This empirically supports
our hypothesis, that our approach can combine
the best of both worlds: very fast learning while
a high accuracy is preserved. In real-world ap-
plications the described approach performs even
better. Applying an automatic feature extraction
method instead of merely applying a feature
selection is very demanding and might take up
to several days while the run-time of the transfer
approach basically remains the same.
In addition to the aforemetioned evaluation,
we visualized the notion of related tasks. Figure
12 shows the base feature vectors of all learning
tasks after a singular value decomposition to
transform the data into two dimensions. On the
one hand, we can see the heterogeneity of learning
tasks. On the other hand, several learning tasks
form groups of tasks with similar base weights.
This approves our hypothesis: while in general
different learning tasks require different features,
some of the learning tasks resemble each other at
least to some extent.