Information Technology Reference
In-Depth Information
Table 5.1.
Dataset description
Dataset
Original size Dims
Classes
Aggregation size Evaluation
Wave
300
21
3
30
leave-1-out
Pima
768
8
2
77
leave-1-out
Shuttle
58000
9
7
594
10-fold
Segment
2310
19
7
319
10-fold
Ringnoise
50000
4
2
500
10-fold
Bank8FM
4499
8 continuous
450
10-fold
Table 5.2.
SVM classification and regression results
Dataset
Accuracy % Mean squared error
Wave
80.00
0.46
Pima
79.22
0.21
Shuttle
94.78
1.10
Segment
91.22
1.70
Ringnoise
84.20
0.19
Table 5.3.
One-class SVM results
Dataset
Outliers
Shuttle
9
Bank8FM
6
classification, regression and novelty detection tasks. There is no experimental
result with interval data mining provided by other algorithms. Therefore, we
only report results obtained by our approach, it is dicult to compare with the
other ones.
5.3.4
Other Kernel-Based Methods
Many multivariate statistics algorithms based on generalized eigenproblems can
also be kernelized [21], e.g. Kernel Fisher's Discriminant Analysis (KFDA), Ker-
nel Principal Component Analysis (KPCA) or Kernel Partial Least Squares
(KPLS). These kernel-based methods can also use the RBF kernel function de-
scribed in section 2 to deal with interval data. We use KPCA and KFDA to
visualize datasets in the embedding space where the user can intuitively see the
separating boundary between the classes based on human pattern recognition
capabilities.
The eigenvectors of the data can be used to detect directions of maximum
variance, and thus, linear PCA is to project data onto principal components by
solving an eigenproblem. By using a kernel function instead of the linear inner
