Applications in Intelligent Music Analysis - Intelligent Audio Analysis - page 271

Table 11.22 summarises all feature subsets as were presented above. A subset
'No-Lyrics' will be used in the ongoing as well, to compare the influence of lyrics
processing in comparison to the information that comes directly from the audio and
artist, title, and year information of a song. It has to be noted at this point that 25 %
(675) of the pieces in the NTWICM database have no lyrics included as these were
not contained in the two used on-line lyric databases—they were left as they are
(empty BoW vector), owing to the philosophy to stick with realism as given for a
working system in a typical use-case.

11.7.2 Performance

Given the imbalance of instances across classes in the NTWICM database (cf.
Sect. 5.3.2 balancing is reasonable to avoid a bias towards class throughout clas-
sifier learning. This was realised by random up-sampling up to perfect balance with
the default random seed in Weka [ 122 ] (cf. Sect. 7.5.1 ) . This required a target size of
200 %.
As classifier serve SMO-trained SVMs with pairwise multi-class discrimination,
linear kernel, and a complexity c of 1 0 at first. The complexity was optimised

on the development set of NTWICM on the A3 and V3 tasks (cf. Sect. 5.3.2 ) and
c

.

. Higher order polynomial kernels did not lead to an
improvement in terms of UA and WA. For reference, performance by RFs will also
be shown.
All results are provided by UA and WA.
First, a feature selection from the 691 overall features was carried out to reveal
promising features and reduce the complexity for the classifier. For easily inter-
pretable feature analysis results, the groups as shown in Table 11.22 are evaluated
individually by classification with the target classifier. An even more compact, but
less interpretable representation in the feature space is then additionally reached by
SFFS—also with the target classifier 'in the loop'. The gold standard throughout fea-
ture selection was given by the rounded median. Table 11.23 summarises the results
of these computations, and Figs. 11.18 and 11.19 visualise the confusions made by
the classifier per feature type.

∈{ 0 . 5 , 1 . 0 , 1 . 5 , 2 . 0 , 2 . 5 }

Table 11.22 Feature subsets

Name Description #

used

Spectral For spectral features 24

Rhythmic For rhythmic features 87

Chords Recognised chord features 22

Meta-Info Date, artist, and title related 153

Lyrics-CN ConceptNet's mood on lyrics 12

Lyrics-BoW Word occurrences in the lyrics 393

All Unision of the above 691

No-Lyrics All without Lyrics-BoW and Lyrics-CN 286