~Proceedings ICMCISMCI2014 14-20 September 2014, Athens, Greece
Table 3. Estimated Results vs. Actual Results
Category Estimated Actual
Radio 5288 10057
Piano 798 752
Drums 644 741
Karaoke 740 246
DAW 226 138
MPC / Sampler 220 136
Table 4. Clustering Results
Cluster Number of Apps Number of Categories
1 6 1
2 94 9
3 29 3
4 191 12
5 3 1
6 14 1
7 27 6
8 57 7
9 24 2
10 63 5
11 19 6
12 2 1
categories of test data was ineffective, even when using the
whitelisted name and the whitelisted description. The apps
both failed to cluster in groups around their categories, and
failed to give correct numbers of apps per cluster. Table 4
shows the number of apps per cluster, and Table 5 shows
the categories per cluster. Figures 1 and 2 shows the results of this clustering, with its dimensionality reduced via
principle component analysis (PCA). As can be seen, each
cluster does not contain only a single category. It was also
hoped that PCA might allow for manual segmentation of
each category. However, as can be seen by the PCA of the
data in Figure 3, this was not possible: the categories are
too intermingled to be able to draw useful segment boundaries.
O ' ke
10-2 tar
Table 5. Clustering Breakdown
Cluster Category Breakdown
1 Radio: 6.
Guitar: 29, Piano: 21, Karaoke: 14, DAW: 9, DJ:
2 8, Amp: 6, Synth: 3, Artist: 2, Sequencer: 2.
3 Drum: 15, MPC: 9, Sequencer: 5.
Artist: 67, Synth: 34, Piano: 25, DJ: 22, Guitar:
4 15, Sequencer: 13, Radio: 5, Amp: 3, MPC: 3,
Drum: 2, DAW: 1, Karaoke: 1.
5 Drum: 3.
6 Radio: 14.
Karaoke 9, Amp: 8, Guitar: 5, Piano: 2, DAW: 2,
Sequencer: 1.
8 Sequencer: 16, DJ 14, Synth: 10, MPC: 10, Drum:
4, DAW: 3.
9 Radio: 23, Artist: 1.
10 Drum: 26, MPC: 19, Sequencer: 13, Synth: 3,
DAW: 2.
11 Amp: 9, DAW: 5, Piano: 2, DJ: 1, Guitar: 1,
Karaoke 1.
12 Radio: 2.
1
4
U -5,sveu~enic r
pequenceq e
eg quet'Weqp.r
UM.
-8
-- -.4
Component I
-2 0
Figure 2. Labeled K-Means clusters, zoomed in.
U.
lii-i o
2.-Iv anoOynth ' t ir t ruh ~
i3 thrtst
pynth yrrth
ptistt '
fth ynth t
}.2 0.4 6 06 10 1.2
Cornponerst I
Irur
15
-20 -10 0 120 30
Component 1
Figure 1. Labeled K-Means clusters.
Figure 3. PCA data, zoomed in.
- 567 -
