Collaborative Use of Features in a Distributed System for the Organization of Music Collections - Intelligent Music Information Systems: Tools and Methodologies - page 162

h ( X ) which predicts the value of Y. Individuals
which provide better classification results when
used as features for the classification task at hand
should have a greater probability to survive into
the next generation. To evaluate the fitness of
each method tree in a population the following
steps are performed:

methods”. Minimizing the number of training
errors and maximizing the margin between the
examples and the hyperplane lead to a constrained
optimization problem, which is solved by intro-
ducing Lagrange multipliers. The resulting op-
timization problem can be efficiently solved, for
example by quadratic programming approaches,
and allows for another interesting property: the
examples are only used in dot products which
might be replaced by other (nonlinear) similarity
functions. The introduction of these nonlinear
kernel functions and the efficiency are the reasons
for the great success of this learning paradigm.
We will later discuss some other properties of
SVM learning in chapter 5 where we exploit
these properties in order to speed up the process
of feature set discovery.

1. Each individual method tree is applied to
an excerpt of the raw data.
2. This method application returns a trans-
formed dataset, which is used by classifier
learning.
3. A k -fold cross validation is executed to
estimate the performances of the learning
scheme in combination with the feature sets
provided by the method trees.
4. The mean accuracy, recall, and/or precision
of the result become the fitness value of the
applied feature construction method trees.
5. The fitness values of the method trees are
used to build the next population with one
of the selection schemes described above.

classifying genres

Since results are published for the genre clas-
sification task, we have applied our approach to
this task, too. Note, however, that no published
benchmark data sets exist. Hence, the comparison
can only show that feature construction and selec-
tion leads to similar performance as achieved by
other approaches. For the classification of genres,
three data sets have been built.

We use a Support Vector Machine (SVM) as
the inner learning scheme. The SVM classifica-
tion is done by calculating an optimal hyperplane
separating the classes. Following the notion of
statistical learning theory, the optimal hyperplane
is the one with the largest safety margin on both
sides (Schölkopf & Smola, 2002). Hence, SVMs
belong to a group of learners called “large margin

• Classic/pop: 100 pieces for each class were

available in Ogg Vorbis format.

Table 2. Classification performance using the
same non-tailored standard feature set for all
classification tasks (linear SVM)

Table 1. Classification of genres with a linear
SVM using the task specific feature sets

Classic/pop Techno/pop Hiphop/pop Classic/pop Techno/pop Hiphop/pop

Accuracy 100% 93 . 12% 82 . 50% Accuracy 96 . 50% 64 . 38% 72 . 08%

Precision 100% 94 . 80% 85 . 27% Precision 94 . 12% 60 . 38% 70 . 41%

Recall 100% 93 . 22% 79 . 41% Recall 95 . 31% 64 . 00% 67 . 65%

Error 0% 6 . 88% 17 . 50% Error 3 . 50% 35 . 63% 27 . 92%