Information Technology Reference
In-Depth Information
Table 4.5 Number of features needed to reach 90% of total accuracy
Rand rank
Goal oriented ranking
OLDA-FR
BVQ-FR
Relief
Gain-ratio
One-rule
HeartStat
9
7
4
11
4
4
Heart
6
5
2
8
3
3
Australian
9
2
4
2
N/D
N/D
Urban
6
3
3
3
3
3
Wildfires
8
4
6
3
5
3
Landslides
8
3
3
6
5
4
Ionosphere
4
4
3
3
4
4
Waveforms
24
6
6
6
8
8
Covertype
9
9
5
5
5
12
Segment
6
4
3
4
3
3
Gottingen
6
3
3
3
6
6
Letter
10
9
7
6
7
7
Corine
16
13
10
12
14
12
Dataset sorted by increasing complexity
Fig. 4.7 Performance Indices cost-benefit of features of EDBFM based ranking. Left Wildfire
dataset. Right CoverType dataset. On the horizontal axis the features sorted by rank and in vertical
the values in percentage normalized to 100%
In the Table 4.6 (top), the significance test is performed on all ranking experiments.
The table does not allowus to assert the superiority of amethod over another; pointing
out that more experiments are needed. However, we have that BVQ-FR overcomes
Relief and Gain Ratio with statistical significance greater than 0.9, while there is
condition of parity with One Rule that is expressed by null statistic significance. It is
noteworthy that BVQ-FR results have been obtained without stressing the parameters
setup of the BVQ algorithm.
 
Search WWH ::




Custom Search