Information Technology Reference
In-Depth Information
requirements are hard to fulfill, is to find those
features of
X
which allow the program to perform
the prediction with high accuracy.
A large variety of features for audio data were
proposed in the literature (e.g., Pampalk, Flexer,
& Widmer, 2005; Tzanetakis, 2002; Tzanetakis
& Cook, 1999 ). However, it has been shown in
several investigations that there does not exist
one set of features which is well suited for all
tasks. For instance, clustering rock music requires
completely different features than does clustering
music according to the mood (Pohle, Pamplak,
& Widmer, 2005). Even in genre classification,
for two different data sets different feature sets
perform better (Lidy & Rauber, 2005). As will
be shown later (Section 4.5), classifying user
preferences requires a different feature set for
each user. Personal taxonomies are far from being
standardized. We cannot even hope to achieve
an appropriate common feature set for the clas-
sification into diverse user taxonomies— even
if we would tailor our feature extractions in a
tremendous effort. The principled problem is that
the particular data set and taxonomy cannot be
foreseen so that there is no standard data set for
which a feature set could be developed. Hence,
we see one opportunity, namely to use machine
learning for the construction of the feature set
itself.
Following this idea, we propose a framework
consisting of two main ideas: first, we systemize
all basic methods of feature extraction and intro-
duce a new concept of constrained combinations,
namely
method trees
. We will then discuss how
these method trees can be automatically found
by means of a genetic programming approach.
Second, we observe that this approach, although
it delivers very accurate feature sets and classi-
fiers, is very time consuming and we propose
a solution for this problem by transferring suc-
cessfully created feature sets to other, similar
learning tasks.
After a more detailed overview of related re-
search, we present a unifying framework which
decomposes the complex feature extractions into
their building blocks (Section 3). This allows us
to automatically learn a feature set for a certain
classification-learning task (Section 4). Since
learning feature sets for learning classifiers is
computationally demanding, we exploit its result
for new, similar classifier learning tasks. Instead
of training the feature extraction anew, we transfer
feature sets to similar classification tasks (Section
5). The new measure of similarity of learning tasks
is interesting also for those who want to manually
tailor feature extraction to a new classification
task, quickly. Finally, in Section 6, we come back
to the user's requirements and illustrate the use
of learning feature extraction and transferring
features by the Nemoz system, a peer-to-peer
music management system.
related research
The automatic organization of music collections
into a hierarchical structure often uses self-or-
ganizing maps (SOMs) (Kohonen, 1988). The
islands of music, for instance, organizes music into
regions of a map taking into account a fixed set of
features (Pampalk, 2001, Pampalk, Dixon, & Wid-
mer, 2004). Due to their pleasant visualization,
SOMs are used to build intelligent user interfaces
to large collections (Rauber, Pampalk, & Merkl,
2002). The burden of structuring a collection is
handed over to the system when applying SOMs.
However, this also means that personal aspects
are not taken into account. A first step towards
personalization is the approach of Baumann and
colleagues (Baumann, Pohle, & Shankar, 2004).
They use feedback of users in order to adjust their
clustering. In addition to audio features they use
lyrics and cultural metadata. Where we share the
goal of hierarchical organization, we rely on audio
features only. Moreover, we want to include the
taxonomies which the users have built. Instead
of replacing the user-made structures, we aim at
enhancing them automatically.
Search WWH ::
Custom Search