Collaborative Use of Features in a Distributed System for the Organization of Music Collections - Intelligent Music Information Systems: Tools and Methodologies - page 165

seems that the variety of genres is reflected in the
variety of extracted features. The classification
task of User 4 again requires a completely different
feature set. For this user, interval features in dif-
ferent spaces were extracted. Another interesting
feature newly discovered for this user was the
slope of a linear regression function in a filtered
frequency space. As a result, we can state that no
feature exists which was extracted for all tasks
and only a few which were extracted for more
than one task. Hence, different user preference
classifications require totally different feature sets
which, in a weaker fashion, was also detected for
genre classifications.

that some of the taxonomies resemble each other
to some extent. This is based on the observa-
tion of sociocultural aspects of music indicat-
ing communities or friend of a friend networks
which share their view of music. We exploit this
fact by transferring successful features among
such similar classification tasks. Thus, instead
of searching for suitable transformations using a
genetic algorithm, we first query other nodes for
features. If relevant features are indeed available,
the given classification task can be learned with
high accuracy and only minimal effort.
This scenario poses some additional con-
straints on the methods used to compare learning
tasks and to share features. First, the retrieval of
similar learning tasks and relevant features has
to be very efficient, as the system is designed for
interactive work. This also means that methods
should enable a best effort strategy, such that the
user can stop the retrieval process at any point
and get the current best result. Second, the system
should scale well with an increasing number of us-
ers and thus learning tasks. Also, it has to deal with
a large variety of heterogeneous learning tasks,
as we cannot make any strict assumptions on the
classification problems users create. Finally, as the
system is distributed, communication cost should
be as low as possible. As a consequence, methods
that are based on exchanging complete data sets
or many feature vectors are not applicable.
We have developed a new similarity measure
of learning tasks from audio data which is based
on already learned feature weights. This measure
is efficient and produces only minimal commu-
nication cost. In this way we obtain specialized
feature sets that achieve high performance in
near real-time.
Let T be the set of all learning tasks, a single
task is denoted by t i . In our scenario each learning
task corresponds to a taxonomy node. Using the
approach discussed in Section 4, we can extract
a feature set X i for task t i . The components of X i
are called features X ik . The objective of every
learning task ti i is to find a function h i ( X i ) which

feature transfer

As already stated, no generic feature set exists
for all possible audio classification tasks. While
learning the feature extraction for a particular
classification task delivers good results, train-
ing the feature extraction is time-consuming
and demands a sufficient set of examples. The
feature extraction process itself takes some time
and for fitness evaluation a full cross-validation
on a learning scheme is performed. For large
classification problems, the whole optimization
process might take up to several days. Therefore,
we discuss an important improvement for the
automatic feature extraction method discussed
above. The basic idea of this improvement is to
check if former work or work of others might be
exploited to reduce the total runtime or to skip
the feature extraction process at all.
Given only one learning task, we cannot do
much about this. In a networked media organizer
system, however, individual nodes are connected
and may share information and data. Most p2p
systems are focused on primarily sharing media
content. In the following we propose an efficient
approach allowing nodes to share features, as
well. Although many different user taxonomies
exist, it is at the same time reasonable to assume