Digital Signal Processing Reference
In-Depth Information
Table 11.23
UA and WA for classification on AllInst test data against different attribute subsets
fortheA3andV3tasks,SVM
Type
Arousal
Valence
[
%
]
UA
WA
UA
WA
Spectral
49.0
47.6
48.8
47.5
Rhythmic
54.0
52.4
57.7
56.4
Chords
50.0
47.0
49.2
47.6
Meta-Info
37.4
36.1
39.3
35.5
Lyrics-CN
33.5
28.9
35.9
38.4
Lyrics-BoW
39.4
36.8
37.8
40.5
All
50.5
50.0
50.9
51.3
No-Lyrics
54.1
53.3
58.8
58.5
(a)
(b)
(c)
(d)
(e)
(f)
Fig. 11.18
Arousal confusions in the A3 classification task for selected feature subsets. Classifier
SVM, dataset AllInst of NTWICM [
28
].
a
Spectral,
b
Rhythmic,
c
Chords,
d
Lyrics-BoW,
e
All,
f
No-Lyrics
As can be seen, the task is demanding, and there are pronounced differences across
individual feature groups: The lyrics features hardly surpass chance level, but the
rhythm features, almost reach the best performance of all features except for lyrics
features. Given this best result for the No-Lyrics set, it will be used in the ongoing.
All features combined being inferior to this reduced set can be seen as indication of a
too high dimensionality of the feature space. Further, the good results for the chord-
based features show the suitability of the 'mid-level' features that base on decisions.
The differences between arousal and valence are less pronounced within a type.
In the confusion matrices for the No-Lyrics and Rhythmic feature sets confusions
luckily occur mostly between neighbouring classes, i.e., negative or positive is mostly
confused with neutral.
