\documentclass[twoside,11pt]{article}
%\documentclass[a4paper]{article}

%\bibliographystyle{plainnat}
%\bibliographystyle{apalike}
\input{psfig.sty}
\usepackage{jmlr2e}

%\jmlrheading{}{2001}{}{}{}{Larry M. Manevitz and Malik Yousef}
\jmlrheading{2}{2001}{139-154}{3/01}{12/01}{Larry M. Manevitz and Malik Yousef}
\ShortHeadings{One-Class SVMs for Document Classification}{Manevitz and Yousef}
\firstpageno{139}

\begin{document}
%\title{One-Class SVM for Document Classification$\ ^{1}$ }
\title{One-Class SVMs for Document Classification}

%\title{One-Class SVM for Document Classification \footnote{This
%work was partially supported by {\it{HIACS}}, the Haifa Interdisciplinary
%Center for Advanced Computer Science.   }}

%\author{\name Larry M. Manevitz \email manevitz@cs.haifa.ac.il \\
\author{\author1 Larry M. Manevitz \email manevitz@cs.haifa.ac.il 
         \AND \author2 Malik Yousef   \email yousef@cs.haifa.ac.il \\
         %\AND
        %\author2 Malik Yousef  \email yousef@cs.haifa.ac.il 
        \addr  Department of Computer Science \\
        University of Haifa \\
        Haifa 31905  Israel}
        %\AND
        %\name Malik Yousef  \email yousef@cs.haifa.ac.il \\
        %\addr  Department of Computer Science \\
        %University of Haifa \\
        %Haifa 31905  Israel }

\editor{Nello Cristianini, John Shawe-Taylor and Bob Williamson}
%\footnotetext[1]{This
%work was partially supported by {\it{HIACS}}, the Haifa Interdisciplinary
%Center for Advanced Computer Science.}

%\author{Larry M. Manevitz and Malik Yousef
%        \\ Department of Computer Science
% \\ University of Haifa \\ Haifa, Israel \\ \ 
%        \\  manevitz@cs.haifa.ac.il  ~ yousef@cs.haifa.ac.il}

\maketitle

\input{0.abstract.tex}

\begin{keywords}
   Support Vector Machine, SVM, Neural Network,
   Compression Neural Network, Text Retrieval,
   Positive Information
\end{keywords}

\input{1.introduction.tex}

\input{1a.Dataandrepresentation.tex}

\section{SVM for One Class Classification: Separating Data from Origin}

The SVM algorithm as it is usually construed is essentially a two-class
algorithm (i.e. one needs negative as well as positive examples).
Below we present two modifications, one due to Sch\"{o}lkopf 
and one we proposed,
to allow its use for only positive data.   Both mechanisms identify
``outliers" amongst the positive examples and use them as negative examples.
Below we refer to the Sch\"{o}lkopf method as ``one-class" and to the other
as ``outlier".

We investigated both algorithms under a variety of choices of parameters.
Our experiments were more extensive for the one-class SVM which seems
to be the superior of the two.



\subsection{Sch\"{o}lkopf Methodology}
%\input{2.SVMoneclass.tex}
\input{2NEW.SVMoneclass.tex}
%TO BE FIXED UP; i.e. the English checked

%MALIK will describe

\subsection{Outlier Methodology}
\input{3.SVMoutlier.tex}
%LARRY and MALIK Needs to be written better

\section{Other Methods}
\input{4.OtherMethods.tex}
%NEED TO EDIT REDUNDANT SECTIONS AT END


%\section{Data Tables}
%\input{6.Tables.tex}
%Scholkopf using 10,20,30 features (10 is best.)
%Scholkopf using Binary, Tf-idf, frequency, Hadamard (with 20 features)

%Scholkopf using Kernels: 


\section{Results}
\input{7.Results.tex}

%\subsection{Scholkopf Results}
%
%Parameters for Scholkopf:
%1.Scholkopf using 10 features is best for SVM.
%2.Scholkopf using Binary is best (for 20 features) (and Malik claims
%%doesnt matter if 10 or 20 or whatever.)  Moreover, VERY BAD using
%%%Tf-idf or Hadamard.   Need explanation as to why so sensitive to 
%representation. 
%3. Scholkopf VERY sensitive to Kernel choice, even within BINARY.
%Moreover, changed between 20 and others for Radial basis is very 
%dramatic.    No explanation.   Malik is double checking the results.
%%
%FINAL CHOICE:  Scholkopf with 10,or 20 but use linear probably, definitely
%with BINARY.
%
%\subsection{Outlier Results}
%Parameters for Outliers
%1. Used best Hamming 7/20.  
%2. Polynomial poor; other kernels compable.
%
%
%NEED TO CHECK:  Can improve outlier by automatically choosing where
%the F1 is optimal.
%
%Sholkopf seems superior to outliers for optimal choices.
%
%
%\subsection{Neural Network Results}
%
%\subsection{Other Methods Results}

%Parameters and Results for Nearest Neighbor:




%Parameters and Results for Rochio:


%Bayes-Parameters




%Parameters and Results for Neural Networks:




%\section{Overall Comparison}

%Looking over the global comparison table, we see that 
%NNs and SVM-oneclass were preferable.   SVM is somewhat easier
%to implement because the optimization of parameters is more direct.
%
%However, the NN results were more robust.
%
%
%Comparison subsection





%\section{Discussion and Future Questions}

%\input{8.SummaryandFutureQuestions.tex}



\acks{This work was partially supported by {\em HIACS}, the Haifa Interdisciplinary Center for Advanced Computer Science.}
\\
\bibliography{bibliography}
\end{document}