{krr,svm,any}
indicates whether the algorithm is designed to optimize the kernel ridge regression (KRR) or support vector machine (SVM) objective.
command\KL algorithm | unif | corr | lin1 | lin2 | quadl2 | align | alignf |
---|---|---|---|---|---|---|---|
klweightfeatures | any | any | krr | krr | |||
klcombinekernels | any | svm | krr | krr | any | any | |
klcombinefeatures | any | svm |
Note, any data can be formatted to work with klcombinekernels
, the programs klcombinefeatures
and klweightfeatures
however give more efficient implementations of algorithms and allow for more efficient representations of data when possible.
The kernel learning algorithms are summarized as follows:
unif
- A uniform linear combination of base kernels/features, regularization restricts the trace of the kernel matrix.
corr
- Weight each feature proportional to its correlation with the labels, regularization restricts the trace of the kernel matrix.
lin1
- A positive linear combination of kernels, regularization restricts the kernel matrix trace. (Lanckriet et al. JMLR 2004, Cortes et al. MLSP 2008)
lin2
- A positive linear combination of kernels, regularization restricts the l2-norm of the weights. (Cortes et al. UAI 2009)
quadl2
- A positive quadratic combination of kernels, regularization restricts the l2-norm of the weights (Cortes et al. NIPS 2009).
align
- A positive linear combination of kernels, with the weight of each kernel chosen proportional to its centered kernel target alignment (Cortes et al. ICML 2010).
alignf
- A positive linear combination of kernel, with the weight vector chosen in order to maximize the kernel target alignment of the final combined kernel (Cortes et al. ICML 2010).
-- AfshinRostamizadeh - 10 Sep 2009