ISIDA/Kernel – User Guide

Author: G. Marcou (University of Strasbourg - 2012)

---

Introduction

ISIDAKernel is a compilation of various kernel based methods. Those methods can be used to compute a Gram matrix, express a covariance matrix in the feature space, extract principal or important components. These methods are typicaly used to derive a "picture" of a dataset or to understand the characterisitcs of a feature space and relations with a particular property.

Command Line Usage

Syntax

Usage: ISIDAKernel [-h|--help] -i <input> -o <output> -m <metric> [-a <alpha>]
[--pairs [-t <threshold>]]
[--graph -t <threshold> [-s <shapefile>] [-l] [-p <proj>]]
[--gram [-w] [-g|-c|-e|p <proj>]]
[--smola [-w] [-n <comp. number>] [-p <proj>]]
[--shawn [-w] [-n <comp. number>] [-p <proj>]]

Options

Option	Description
-i	input file in SVM format. Each line represent an object. The first column must be a numerical property. Each other word of the line is of the form "a:b" where "a" is an integer, the index of a non-null component of the vector describing the object and "b" is the actual value of the component.
-m	Select a kernel
	2: Tanimoto kernel $(Tc ^alpha)$
	3: RBF $(exp(-alpha||u-v||^2)$
	4: Cosinus $[(u*v/(||u||.||v||)+1)/2] ^alpha$
	1: euclidean distance $||u-v||$. This is not a kernel.
-a	value of a parameter of a kernel. Default value is 1.
-w	The first column of the SVM file are interpreted as weights. Kernels are bilinear forms of those weights. If the object u and v have wheights w_u and w_v then K(w_u.u, w_v.v)=w_u.w_v.K(u,v). This is a definition of weights, whatever the mathematical form of a kernel is.
-t	threshold value used in some methods. Default value is 0.
-n	number of component to extract used in some methods
-p	Project new data on an existing representation.
--pairs	compute for each pair of object the kernel. Outputs only those pairs that are more similar than a thershold.
--graph	compute a multicomponent graph in which each node is an object and two objects are connected by an edge if they are more similar than a given threshold. Outputs a graph in GraphML format that can be easily vizualized and manipulated using a software such as Yed. The method understand some display command for the nodes.
-s	file indicating for each node the color and the shape used to represent it. One line per node. For instance "orange square" on line 12 will represent the node 12 by an orange square.
-l	Use Laplace Eigenvalues of each components. For each component outputs the Laplace Eigenvalues and Eigenvectors. The two first component are used to produce a layout for each component.
--gram	Compute the Gram matrix or the Covariance matrix or the Kernel PCA.
-g	Compute the Gram matrix.
-c	Compute the covariance matrix.
-e	Perform the Kernel Principal Component Analysis.
--smola	Perform a Kernel Principal Feature Analysis using Smola's algorithm.
--shawn	Perform a Kernel Principal Feature Analysis using Shawn's algorithm.