Multiclass dataset

We illustrate the application of support vector machines for aroma classification using as our example 98 tetra-substituted pyrazines (Figure 52) representing three odor classes, namely 32 green, 23 nutty, and 43 bell-pepper. The prediction power of each SVM model was evaluated with a leave-10%-out cross-validation procedure. This multiclass dataset was modeled with an one-versus-all approach. 

Figure 7. Multiclass NaTve Bayes modeling within Pipeline Pilot software (www.scitegic.com) based on the WOMBAT chemogenomics dataset Probabilistic target predictions are possible for compounds given only their chemical structure. In the example shown, the WOMBAT targets were predicted for Calphostin C, a known protein kinase C inhibiting natural product Tubulin and beta-hexosaminidase are predicted as additional possible targets.

The one-versus-all procedure requires a much smaller number of models, namely for a k-class problem, only k SVM classifiers are needed. The ith SVM classifier is trained with all patterns from the ith class labeled - -1, and all other patterns labeled 1. Although it is easier to implement than the one-versus-one approach, the training sets may be imbalanced due to the large number of —1 patterns. In a comparative evaluation of one-versus-one, one-versus-all, and DAGSVM methods for 10 classification problems, Hsu and Lin found that one-versus-all is less suitable than the other methods." However, not all literature reports agree with this finding. Based on a critical review of the existing literature on multiclass SVM and experiments with many datasets, Rifkin and Klautau concluded that the one-versus-all SVM classification is as accurate as any other multiclass approach. ... 

The SVM classification of vertical and horizontal two-phase flow regimes in pipes was investigated by Trafalis, Oladunni, and Papavassiliou. " The vertical flow dataset, with 424 cases, had three classes, whereas the horizontal flow dataset, with 2272 cases, had five classes. One-versus-one multiclass SVM models were developed with polynomial kernels (degrees 1 to 4). The transition region is determined with respect to pipe diameter, superficial gas velocity, and superficial liquid velocity. Compared with experimental observations, the predictions of the SVM model were, in most cases, superior to those obtained from other types of theoretical models. 

---