\documentclass[a4paper]{article}
\usepackage{hyperref, graphicx, color, alltt, amsmath}
\usepackage{Sweave}
\usepackage[round]{natbib}

\begin{document}
\SweaveOpts{concordance=TRUE}

%\VignetteIndexEntry{Weighted Support Vector Machine Formulation}
%\VignetteDepends{WeightSVM}
%\VignetteKeywords{classification, regression, machine learning, support vector machines}
%\VignettePackage{WeightSVM}

\SweaveOpts{engine=R,eps=FALSE}
\setkeys{Gin}{width=0.8\textwidth}

\title{Weighted Support Vector Machine Formulation
\author{by Tianchen Xu}
\url{tx2155@columbia.edu}
}
\maketitle
\sloppy

The original formulation of unweighted SVM with linear kernel is as follows \cite{valdimirnature}:
  \begin{eqnarray}
    \min_{\omega,\xi}&&\frac{1}{2}\|\omega\|^2+C \sum_{i=1}^n (\xi_i+\xi_i^*) \nonumber\\
    \mbox{s.t.} &&y_i - \langle\omega,x_i\rangle-\omega_0\le \varepsilon + \xi_i,\nonumber\\
    &&\langle\omega,x_i\rangle+\omega_0-y_i \le \varepsilon + \xi_i^*,\nonumber\\
    &&\xi_i, \xi_i^* \ge 0.\nonumber
\end{eqnarray}
The constant $C>0$ determines the trade-off between the flatness of $f$ and the amount up to which deviations larger than $\varepsilon$ are tolerated. This corresponds to dealing with a so called
$\varepsilon$-insensitive loss function $|\xi|_\varepsilon$ described by
\begin{equation}
|\xi|_\varepsilon =
\begin{cases}
  0, & \text{if }|\xi|\le \varepsilon\\
  |\xi|-\varepsilon, & o/w.
\end{cases}\nonumber
\end{equation}\\



The corresponding weighted SVM with $W_i$ as individual weights:
  \begin{eqnarray}
    \min_{\omega,\xi}&&\frac{1}{2}\|\omega\|^2+C \sum_{i=1}^n \textcolor{red}{W_i}(\xi_i+\xi_i^*) \nonumber\\
    \mbox{s.t.} &&y_i - \langle\omega,x_i\rangle-\omega_0\le \varepsilon + \xi_i,\nonumber\\
    &&\langle\omega,x_i\rangle+\omega_0-y_i \le \varepsilon + \xi_i^*,\nonumber\\
    &&\xi_i, \xi_i^* \ge 0.\nonumber
\end{eqnarray}\\

Other kinds of weighted SVMs (with different kernels) have the similar formuation.\\
\noindent Available kernels:\\
\\
\noindent
\begin{minipage}{\textwidth}
  \centering
  \begin{tabular}{|l|l|l|} \hline
    kernel            & formula & parameters \\ \hline \hline
    linear            & $\bf u^\top v$& (none) \\
    polynomial        & $(\gamma \mathbf{u^\top v}+c_0)^d$ & $\gamma, d, c_0$\\
    radial basis fct. & $\exp\{-\gamma|\mathbf{u-v}|^2\}$&$\gamma$\\
    sigmoid           & $\tanh\{\gamma \mathbf{u^\top v}+c_0\}$ &$\gamma, c_0$\\ \hline
  \end{tabular}
\end{minipage}

\begin{thebibliography}{1}

\bibitem[Valdimir and Vapnik(1995)]{valdimirnature}
V~Valdimir and N~Vapnik.
\newblock The nature of statistical learning theory.
\newblock 1995.

\end{thebibliography}

\end{document}