Learning Kernels Documentation

Command Line Binaries

There are three main command line binaries that can be used for generating a kernel or feature mapping, depending on how the base matrices are defined. They are listed here in order from the most to least general.

• klcombinekernels - given several kernel matrices for the same dataset, output a weighted combination of these kernel matrices.
• klcombinefeatures - given several explicit feature mappings of the same dataset, output a weighted combination of these feature mappings.
• klweightfeatures - given a single explicit feature mapping of the dataset, output a weighted feature mapping or kernel. I.e. each feature in this case corresponds to a rank-1 kernel.

Usage

$ command  [flags] input_file output_file

The input_file is either an explicit feature mapping representation of the data or a kernel matrix representation, both of which should follow the LIBSVM format. Additionally, klcombinekernels can read binary kernel files output by the kleval tool with the --bin flag set.

The output_file is also going to be an explicit feature mapping representation or kernel matrix representation of the data in LIBSVM format.

Command Line Flags

Here we give a list of useful flags that are used in conjunction with the command line binaries. Note, running any of the commands without any arguments will result in a full listing and description of possible flags and their default values.

General Flags

• --alg_reg - For algorithm specific kernel combination methods, use the specified algorithm regularization parameter.
• --ker_reg - Specify kernel regularization parameter use within the kernel combination algorithm.
• --interp_param - The interpolation parameter used with iterative kernel combination algorithms, chosen between 0 and 1. A value closer to 0 will lead to larger steps, but also possible instability.
• --lk_alg - Select which kernel combination method is used. See feature table to see which combination methods are available with each command line binary.
• --max_iter - The maximum number of iterations used by any iterative kernel combination algorithm.
• --mu_file - Save the kernel combination weights to the specified file.
• --num_train - Give argument p:q to select points p through q for training the matrix. The learned kernel function is still applied to the entire dataset.
• --sparse - Used sparse data-structures when given feature mappings as input.
• --tol - Specify tolerance of stopping criteria for iterative kernel combination algorithms. Smaller values lead to more precise answers, but longer convergence times.

Command Specific Flags

• --read_kar - expect the input to be a binary kernel matrix.
• --label_file - this flag must be set when the --read_kar flag is used. Should point to the location of a text file with one label per line, which corresponds to the binary matrix input.

klweightfeatures:

• --features - print output using explicit feature mappings instead of computing the kernel.
• --offset - include the specified constant value as an additional feature (can be used to act as an offset).

-- AfshinRostamizadeh - 10 Sep 2009