Information Technology Reference
In-Depth Information
•
Users can search for music similar to a
selected song.
conclusIon
•
A taxonomy can be enhanced through the
taxonomies of other users automatically
(distributed collaborative clustering (Wurst
& Morik, 2006)).
In this chapter we analyzed the problem of
supervised feature extraction and transfer for
organizing personal music collections. Audio
classification enables users to maintain arbitrary
personal organization schemes easily. New audio
files are assigned automatically to the user speci-
fied classes, such that the effort for the user is
minimal. While this approach is very attractive,
one problem needs to be solved in order to apply
automatic classification techniques to audio files.
Audio classification needs an adequate feature
representation of the given data set. Which
features are well suited strongly depends on the
classification task at hand.
We described a unifying framework which
decomposes the complex extraction methods into
their building blocks. This systematization allows
us to move beyond the manual composition of
feature extraction methods for a given classifica-
tion task. This way, the design of an appropriate
feature set has become a learning task in its own
right. Our genetic learning approach discovered
several of the well-known features as well as
some new ones. The automatic feature extraction
approach delivers excellent results in terms of
accuracy for a wide range of different learning
tasks and is able to replace the phase of manual
feature tuning in future applications.
A drawback of this and similar methods is,
however, that they are computationally very de-
manding. In addition, it is required that the user
classifies a considerably large amount of audio
files. We therefore embedded the new learning
task of feature extraction in a (distributed) meta-
learning setting. We exploit the fact that there is
not only one, but several audio classification tasks
faced by many different users or by the same user
in different points of time. The idea is to exploit
the former, time-consuming runs of feature
extraction learning and transfer the resulting
feature sets to similar new classification tasks.
We hence allow users to share features among
Which methods are used in a particular func-
tion can easily be configured through the learning
component provided by the Y
ale
system. Cur-
rently, we apply hierarchical single-link cluster-
ing for the automatic construction of taxonomies
using a fixed set of audio features. For classifier
learning we used instance-based learning, deci-
sion-tree learning, and the support vector machine
together with forward feature selection.
Nemoz is implemented as a p2p network and
allows nodes to share arbitrary information. A
major focus of Nemoz is on sharing features that
are used to learn classification functions. A node
uses the procedure described in Section 5 to query
other nodes for features given a particular clas-
sification task. When a node receives responses
to its query, it applies the feature construction
procedures in the response set to extract the cor-
responding features from the local audio data and
then repeats the learning process. All successfull
features are stored for future use and to respond
to the queries of other nodes.
Nemoz supports still another form of feature
sharing, based on extensional queries. A node
queries other nodes with a set of IDs representing
audio files. Other nodes then respond with feature
values directly instead of feature construction
procedures. This approach is more expensive in
terms of communication costs, can however be
applied even to nominal features like label infor-
mation or ID3 tags (Wurst & Morik, 2006).
Beside feature sharing, we exploit the p2p
structure for other kinds of queries as well, such
as search and distributed clustering. Nemoz is
designed to be extendible, that is, additional
functionality and communication patterns can
be added easily.
Search WWH ::
Custom Search