Information Technology Reference
In-Depth Information
Table 4.4 The EDBFM ranking is compared to other methods (Relief, Gain Ratio, OneRule); the
Performance Index (
) is calculated for each of the accuracy curves in Fig. 4.5
Goal oriented ranking
OLDA-FR (
ˆ
ˆ
)
BVQ-FR (
ˆ
)
Relief (
ˆ
)
Gain-ratio (
ˆ
)
One-rule (
ˆ
)
HeartStat
0.116
0.411
0.059
0.516
0.513
Heart
0.187
0.471
0.151
0.548
0.521
Australian
0.685
0.463
0.602
0.744
0.806
Urban
0.543
0.465
0.408
0.331
0.609
Wildfires
0.474
0.354
0.483
0.336
0.647
Landslides
0.400
0.388
0.292
0.251
0.402
0.096 0.207 0.083 0.075 0.072
Waveform 0.651 0.670 0.654 0.614 0.621
CoverType 0.125 0.373 0.418 0.264 0.040
Segment 0.348 0.595 0.483 0.501 0.467
Gottigen 0.445 0.456 0.407 0.169 0.187
Letter 0.133 0.287 0.349 0.198 0.217
Corine 0.484 0.592 0.438 0.474 0.545
Mean 0.346 0.441 0.371 0.386 0.434
Variance 0.055 0.016 0.034 0.039 0.056
The two bottom rows are descriptive statistics of the Performance Index computed values
Ionosphere
Fig. 4.6 Each piecewise line represents a method of ranking, each radial line represents a dataset.
Datasets are radially ordered by increasing complexity. The intersections represent the values of
performance index
ˆ
to support the observations made. The probability of success of an algorithm over
another is calculated with the binomial distribution , from the count of victories and
defeats, or the number of times that the algorithm outperformed the others on the
basis of the index of performance. Assuming the null hypothesis is that the frequency
of success of the two algorithms is the same, with the two-tailed test it can be seen
how much we deviate from “null hypothesis” assumption.
 
Search WWH ::




Custom Search