Information Technology Reference
In-Depth Information
user preferences
taken from the same genre is not supported
(User 2 ). Surprisingly good is the learning result
for a broad variety of genres among the favorites
(User 3 ). This fact indicates that for this user the
constructed feature set supports the building of
preference clusters in feature space instead of
dominating genre clusters. In contrast to this
result the (negative) effect of sampling from few
records can be seen clearly (User 4 ). Applying the
learned decision function to a database of records
allowed the users to assess the recommendations.
They were found very reasonable. Users agree that
no music which was disliked was recommended
to the user, but unknown plays and those, which
could have been selected as the top 50.
It has already been shown in several investiga-
tions that there does not exist one set of features
which is well suited for all tasks (Lidy & Rau-
ber, 2005; Pohle et al., 2005). The same applies
in our case where automatic feature extraction
finds optimal feature sets by means of genetic
programming. The difference between feature
sets found for different users seems to be even
greater compared to those found for genre clas-
sification. The features most often used for User 1
are those modeling timbre distance, that is both
the peaks in the spectrum of the given audio files
and measurements like the quotient of minimum
or maximum and average (like spectral crest
factor). This is reasonable since the differences
between the positive and negative classes can be
explained by different sound characteristics. For
User 2 , very simple features were extracted like
the length of the songs and the root mean square
average and variance which measures the loudness
and its variance. In addition, some features in time
space were extracted, namely zero crossing rate
and autocorrelation. For User 3 a broad range of
very different features were extracted, including
spectral features similar to those extracted for
the classification task posed by the first user. In
addition, the differences and variances between
the time space extrema together with features in
phase space were also extracted for this user. It
Recommendations of songs to possible customers
are currently based on the individual correlation of
record sales. This collaborative filtering approach
ignores the content of the music. A high correlation
is only achieved within genres, because the prefer-
ences traversing a type of music are less frequent.
The combination of favorite songs into a set is a
very individual and rare classification. It is not a
generalization of many instances. Therefore, the
classification of user preferences beyond genres
is a challenging task, where for each user the fea-
ture set has to be learned. Of course, sometimes
a user is interested only in pieces of a particular
genre. This does not decrease the difficulty of
the classification task. In contrast, if positive and
negative examples stem from the same genre, it is
hard to construct distinguishing features. Genre
characteristics might dominate the user-specific
features. As has been seen in the difficulty of the
data set for hiphop vs. pop, sampling from few
records also increases the difficulty of learning.
Hence, four learning tasks of increasing difficulty
have been investigated.
Four users brought 50 to 80 pieces of their
favorite music ranging through diverse genres.
They also selected the same number of negative
examples. User 1 selected positive examples from
rock music with a dominating electric guitar. User
2 selected positive as well as negative examples
from jazz music. User 3 selected music from a
large set of different genres containing classic,
Latin, soul, rock, and jazz. User 4 selected pieces
from different genres but only from few records.
Using a 10-fold cross validation, mySVM was
applied to the constructed and selected features,
one feature set per learning task (user). Table 4
shows the results.
The excellent learning result for a set of posi-
tive instances which are all from a certain style
of music corresponds to our expectation (User 1 ).
The expectation that learning performance would
decrease if positive and negative examples are
Search WWH ::




Custom Search