Collaborative Use of Features in a Distributed System for the Organization of Music Collections - Intelligent Music Information Systems: Tools and Methodologies - page 151

Feature extraction for audio classification is the
dominant issue investigated in the proceedings of
conferences and workshops about music informa-
tion retrieval. Since 1998 (Liu, Wang, & Chen,
1998; Tzanetakis, 2002; Zhang & Kuo, 1998), an
ever increasing set of feature extractions has been
developed. There are still new features proposed
even for the genre classification task (Pampalk
et al., 2005). We expect that there will never be
the complete and appropriate feature set for all
tasks. Hence, we do not participate in the effort
of constructing new features manually, but have
machine learning methods constructing features
for us. Funny enough, our first publication on
automatic extraction of audio features (Mierswa
& Morik, 2005) was not perceived as it was meant
to. Readers cited this publication as the presen-
tation of a new feature type, namely features in
the phase space. This state space reconstruction
is included in our general framework of building
blocks for feature construction in order to com-
plete the general basis transformations. In fact, it
delivers good results sometimes. However, our aim
is far more challenging, namely to automatically
adapt feature construction to a classification task
at hand. In the following, we will describe this
automatic process, where phase space features
are just another building block which might be
combined with other building blocks.
Computational support in constructing fea-
tures usually means a library of mathematical
functions, for instance using M atlab (Pampalk
et al., 2004). System support is also included in
the M arsyas system, but it is fixed to 30 feature
extractions in version 1.0 and 80 feature extrac-
tions in version 2.0 (Bray & Tzanetakis, 2005;
Tzanetakis & Cook, 1999, ). Also libraries of
some known feature extractions are made pub-
licly available, for example, in the ACE and the
jAudio system (McEnnis, McKay, Fujinaga, &
Depalle, 2005; McKay, Fiebrink, McEnnis, Li,
& Fujinaga, 2005). The Y ale environment with
its audio and value series plug-in moves beyond
this library approach 1 (Fischer, Klinkenberg,

Figure 1.The machine learning environment Y ale
includes a plugin containing all described build-
ing blocks and methods for feature extraction
together with operators for machine learning and
model evaluation

Mierswa, & Ritthoff, 2002). It does not publish
a set of complex feature extractions, but offers a
set of building blocks for feature construction (cf.
Section 3). A large variety of machine learning
algorithms including all W eka algorithms is pro-
vided to users. It allows users to run experiments
which automatically construct a feature set for a
certain learning task and performance measure
(Section 4). The Y ale user chooses which clas-
sification learner the user wants to couple with the
automatic feature extraction in the experiment.
Since the output is in standard representation, it
cannot only be used by the Y ale environment, but
also by other learning environments. Hence, the
support in running a learning experiment does
not prescribe a particular classification learner.
In contrast, the user is in complete control of
designing the experiment (see Figure 1).
Our claim that feature extraction needs to
be adapted to a particular classification task and
dataset is supported by the extensive comparison
of seven feature sets in Mörchen, Ultsch, Nöcker,
and Stamm (2005). They eagerly enumerate