Information Technology Reference
In-Depth Information
nemoz
(performer, composer, album, year, duration of
the song, genre, and comment) and a reference to
the location of the song, but also features which
are extracted from the raw data. The feature
extraction is the application of the results of the
automatic feature construction (Section 4). Since
the learning experiment delivers different results
for different learning tasks, the set of features
may vary. Developers can, however, use a fixed
set of features, if running a learning experiment
is too time consuming.
For intelligent functions, the data structure of
taxonomies is implemented. A collection can be
organized using several taxonomies in parallel.
For an example, see Figure 14. At each taxonomy
node, a classification function can be stored,
which decides whether a new song belongs to this
node, or not. This classifier is the learned model,
where the learning task has been performed by
one of the methods supplied by Yale. Based on
this data structure Nemoz already implements
several intelligent functions:
Together with a group of students we have de-
veloped Nemoz as a framework for studying
collaborative music organization
3
. It has a graphi-
cal user interface with several visualizations, an
application layer with generic operations which
offer various ways to exploit the services which,
in turn, access the data model, and a developer's
interface for fast modification of system behav-
ior. The architecture of the system is shown in
Figure 13.
Nemoz is implemented such that it can be
tailored to:
•
Different learning tasks
•
Different media
•
Different network architectures
Introducing new learning tasks is easy because
of the connection (LabService) to the Y
ale
system
(Fischer et al., 2002), in which diverse classifica-
tion, clustering, and preprocessing operators can
be composed to form an experiment. Storing other
than music data is easy because the Descrip-
torService calls feature extraction methods, which
result in complete or stripped sets of attributes
for given raw data. The NetworkService currently
implements communication among LAN nodes
via TCP and UDP. In addition to tailoring the
system behavior using the implemented func-
tions, the abstract classes and interfaces can be
implemented in order to enlarge the scope of the
system. Moreover, the PluginService together with
the ScriptEngine allows introduction of new codes
and connects Nemoz to other systems.
Nemoz is made for experimenting with intelli-
gent functionality of media organization systems.
Of course, the basic functionality is implemented:
download and import of songs, playing music,
retrieving music from a collection based on
given metadata, and creating play lists. The data
model covers not only the standard metadata
•
The taxonomies can be defined extensionally
by the user, or
•
Automatically be learned using hierarchical
clustering.
•
Taxonomy nodes can be intensionally de-
fined by a (learned) model.
•
New music can be classified into the tax-
onomy.
•
Users can view the audio collections of
other users in terms of an own classification
scheme by temporarily classifying the con-
tent of this collection into an own taxonomy
(goggling).
•
Search queries can be composed using
AND and OR, quantor-free logic formulas
without negation allowing queries such as
“users who are less than 18 years old and
who own a taxonomy with the word grunge
in a node's title”.
•
Users can search for similar taxonomies in
the network.
Search WWH ::
Custom Search