ܟ, ߚ^ଵሻൌ൬2ߨ൰

exp ൜െ2 ሺܡെ઴ܟሻ^௧ሺܡെ઴ܟሻൠ

(3.73)

yesian learning procedure is used for RVM, by which the a priori

is set for the linear coefficients given a vector of the hyper-

rs for w,

݌ሺܟ|ߙሻൌෑܩሺݓ௡|0, ߙ௡^ିଵሻ

ே

௡ୀଵ

(3.74)

osterior over w is derived using the Bayes rule and the marginal

d is established for the model [Tipping, 2001]. RVM has been

or EEG signal classification, where the commonly used Gaussian

nomial kernels were replaced by a chaotic dynamic kernel [Dong,

18]. It was also used for the protein-protein interaction prediction

[Li, et al., 2018].

R package for RVM is implemented in kernlab. The package

a number of different kernel functions. The use of a relevance-

achine based on kernlab package is shown below,

rvm(x,y,kernel,kpar)

e 3.35(a) shows the RVM model constructed for a two-cluster

The relevance vectors were those surrounding two clusters.

arly, the kernels used in RVM can also be replaced by the bio-

ction for modelling peptides data. Doing so, the model is called a

M. A bio-RVM model was constructed for the factor Xa protease

data. In the model, bio-basis function was supported by the

mutation matrix for the homology score calculation between

and relevance vectors. The Jackknife test was used for the

ation test of the model. Figure 3.35(b) shows the ROC curve of

nifed model. The AUC value was 0.962.