Page 132 - Data Science Algorithms in a Week

116 Loris Nanni, Sheryl Brahnam and Alessandra Lumini

Table 4: Performance comparison of some ensembles compared with stand-alone
approaches and previous results.

DATASET MR1 MR2 MR3 (Loris Nanni MR3+(Loris MR4 MR5 1D
et al., 2012) Nanni et al.,
2012)

breast 99.2 99.2 99.4 97.4 99.3 99.3 99.4 99.3
heart 90.2 90.3 89.9 90.1 90.4 90.5 90.5 89.5

pima 80.8 80.9 81.8 71.9 81.3 82.3 82.5 82.4
sonar 94.3 94.3 93.0 92.8 93.2 95.4 95.6 95.2
iono 98.4 98.4 98.2 98.4 98.4 98.3 98.2 98.1
liver 73.9 73.7 74.8 70.3 73.6 76.2 75.8 75.6

hab 67.6 67.8 69.0 59.2 65.8 70.0 69.1 70.1
vote 97.7 97.7 97.7 97.7 97.7 98.5 98.5 98.5
aust 91.7 91.7 92.0 90.8 91.7 92.1 92.4 92.0

trans 67.2 69.5 70.6 61.9 65.8 72.5 73.0 72.9
wdbc 99.5 99.5 99.5 98.8 99.5 99.6 99.6 99.6
bCI 96.1 96.4 96.2 97.0 96.8 96.3 96.4 95.6

pap 86.1 87.2 87.3 88.0 87.5 87.5 87.4 86.8
torn 94.1 94.5 94.6 93.6 94.7 94.2 94.5 90.2
gCr 77.2 78.4 79.7 78.9 79.7 80.7 80.7 80.1

Average 87.6 88.0 88.2 85.8 87.7 88.9 88.9 88.4

This study expands our previous research in this area. First, it investigates different
methods for matrix representation in pattern classification. We found that approaches
based on FFT worked best. Second, we explored the value of using different texture
descriptors to extract a high performing set of features. Finally, we tested the
generalizability of our new approach across several datasets representing different
classification problems. The results of our experiments showed that our methods
outperformed SVMs trained on the original 1D feature sets.
Because each pixel in a texture describes a pattern that is extracted starting from the
original feature, we were also motivated to investigate the correlation among the original
features belonging to a given neighborhood. Thus, we studied the correlation among
different sets of features by extracting images from each pattern and then randomly