\relax 
\bibstyle{plainnat}
\citation{jain}
\citation{fukunaga}
\citation{dhs}
\citation{kmeans1}
\citation{shamir-gen}
\citation{blatt}
\citation{roberts}
\citation{fukunaga}
\citation{Vapnik95}
\citation{sch-sup}
\citation{sch-sup1}
\citation{tax-duin}
\citation{vclust}
\citation{nips00}
\@writefile{toc}{\contentsline {section}{\numberline {1}Introduction}{125}}
\citation{roberts}
\citation{sch-sup}
\citation{tax-duin}
\@writefile{toc}{\contentsline {section}{\numberline {2}The SVC Algorithm}{126}}
\@writefile{toc}{\contentsline {subsection}{\numberline {2.1}Cluster Boundaries}{126}}
\newlabel{outliers}{{1}{126}}
\citation{fletcher}
\citation{tax-duin}
\newlabel{lagrangian}{{2}{127}}
\newlabel{sumbeta}{{3}{127}}
\newlabel{eqn-a}{{4}{127}}
\newlabel{upperbound}{{5}{127}}
\newlabel{kkt1}{{6}{127}}
\newlabel{kkt2}{{7}{127}}
\newlabel{boundbeta}{{9}{127}}
\newlabel{radius}{{13}{128}}
\newlabel{R}{{14}{128}}
\newlabel{contour}{{15}{128}}
\@writefile{toc}{\contentsline {subsection}{\numberline {2.2}Cluster Assignment}{128}}
\@writefile{toc}{\contentsline {section}{\numberline {3}Examples}{128}}
\newlabel{four}{{3.1}{129}}
\@writefile{lof}{\contentsline {figure}{\numberline {1}{\ignorespaces Clustering of a data set containing 183 points using SVC with \ensuremath  {C=1}. Support vectors are designated by small circles, and cluster assignments are represented by different grey scales of the data points. (a): \ensuremath  {q=1} (b): \ensuremath  {q=20} (c): \ensuremath  {q=24} (d): \ensuremath  {q=48}. }}{129}}
\@writefile{toc}{\contentsline {subsection}{\numberline {3.1}Example without BSVs}{129}}
\@writefile{toc}{\contentsline {subsection}{\numberline {3.2}Example with BSVs}{129}}
\citation{sch-sup}
\citation{vclust}
\newlabel{sv-plot}{{3.1}{130}}
\@writefile{lof}{\contentsline {figure}{\numberline {2}{\ignorespaces  Number of SVs as a function of \ensuremath  {q} for the data of Figure 1. Contour splitting points are denoted by vertical lines. }}{130}}
\newlabel{outlier_number}{{17}{130}}
\newlabel{ringnorm}{{3.2}{131}}
\@writefile{lof}{\contentsline {figure}{\numberline {3}{\ignorespaces Clustering with and without BSVs. The inner cluster is composed of 50 points generated from a Gaussian distribution. The two concentric rings contain 150/300 points, generated from a uniform angular distribution and radial Gaussian distribution. (a) The rings cannot be distinguished when \ensuremath  {C=1}. Shown here is \ensuremath  {q=3.5}, the lowest \ensuremath  {q} value that leads to separation of the inner cluster. (b) Outliers allow easy clustering. The parameters are \ensuremath  {p=0.3} and \ensuremath  {q=1.0}. }}{131}}
\newlabel{overlap1}{{3.2}{131}}
\@writefile{lof}{\contentsline {figure}{\numberline {4}{\ignorespaces Clusters with overlapping density functions require the introduction of BSVs. }}{131}}
\citation{dhs}
\citation{roberts}
\citation{ripley}
\citation{ripley}
\citation{ripley}
\@writefile{toc}{\contentsline {section}{\numberline {4}Strongly Overlapping Clusters}{132}}
\newlabel{density}{{22}{132}}
\citation{iris}
\citation{UCI}
\citation{bottleneck-nips}
\citation{blatt}
\newlabel{overlap2}{{4}{133}}
\@writefile{lof}{\contentsline {figure}{\numberline {5}{\ignorespaces In the case of significant overlap between clusters the algorithm identifies clusters according to dense cores, or maxima of the underlying probability distribution. }}{133}}
\newlabel{figure-crabs}{{4}{133}}
\@writefile{lof}{\contentsline {figure}{\numberline {6}{\ignorespaces  Ripley's crab data displayed on a plot of their 2nd and 3rd principal components: (a) Topographic map of \ensuremath  {P_{svc}({\bf  x})} and SVC cluster assignments. Cluster core boundaries are denoted by bold contours; parameters were \ensuremath  {q=4.8, p=0.7}. (b) The Parzen window topographic map \ensuremath  {P_{w}({\bf  x})} for the same $q$ value, and the data represented by the original classification given by \cite  {ripley}. }}{133}}
\citation{jain}
\newlabel{figure-iris}{{4.1}{134}}
\@writefile{lof}{\contentsline {figure}{\numberline {7}{\ignorespaces  Cluster boundaries of the iris data set analyzed in a two-dimensional space spanned by the first two principal components. Parameters used are \ensuremath  {q=4.2\,\, p=0.55}. This resulted in 4 misclassifications. }}{134}}
\@writefile{toc}{\contentsline {subsection}{\numberline {4.1}The Iris Data}{134}}
\citation{milligan}
\citation{psb02}
\citation{SMO}
\citation{sch-sup}
\citation{SMO}
\citation{vclust}
\@writefile{toc}{\contentsline {subsection}{\numberline {4.2}Varying $q$ and $p$}{135}}
\@writefile{toc}{\contentsline {section}{\numberline {5}Complexity}{135}}
\citation{jain}
\citation{lipson}
\bibdata{svc}
\bibcite{psb02}{{1}{2001{a}}{{Ben-Hur et~al.}}{{Ben-Hur, Elisseeff, and Guyon}}}
\bibcite{vclust}{{2}{2000}{{Ben-Hur et~al.}}{{Ben-Hur, Horn, Siegelmann, and Vapnik}}}
\bibcite{nips00}{{3}{2001{b}}{{Ben-Hur et~al.}}{{Ben-Hur, Horn, Siegelmann, and Vapnik}}}
\bibcite{UCI}{{4}{1998}{{Blake and Merz}}{{}}}
\bibcite{blatt}{{5}{1997}{{Blatt et~al.}}{{Blatt, Wiseman, and Domany}}}
\bibcite{dhs}{{6}{2001}{{Duda et~al.}}{{Duda, Hart, and Stork}}}
\bibcite{iris}{{7}{1936}{{Fisher}}{{}}}
\@writefile{toc}{\contentsline {section}{\numberline {6}Discussion}{136}}
\bibcite{fletcher}{{8}{1987}{{Fletcher}}{{}}}
\bibcite{fukunaga}{{9}{1990}{{Fukunaga}}{{}}}
\bibcite{jain}{{10}{1988}{{Jain and Dubes}}{{}}}
\bibcite{lipson}{{11}{2000}{{Lipson and Siegelmann}}{{}}}
\bibcite{kmeans1}{{12}{1965}{{MacQueen}}{{}}}
\bibcite{milligan}{{13}{1985}{{Milligan and Cooper}}{{}}}
\bibcite{SMO}{{14}{1999}{{Platt}}{{}}}
\bibcite{ripley}{{15}{1996}{{Ripley}}{{}}}
\bibcite{roberts}{{16}{1997}{{Roberts}}{{}}}
\bibcite{sch-sup}{{17}{2000}{{Scholkopf et~al.}}{{Scholkopf, Williamson, Smola, Shawe-Taylor, and Platt}}}
\bibcite{sch-sup1}{{18}{2001}{{Scholkopf et~al.}}{{Scholkopf, Platt, Shawe-Taylor, , Smola, and Williamson}}}
\bibcite{shamir-gen}{{19}{2000}{{Shamir and Sharan}}{{}}}
\bibcite{tax-duin}{{20}{1999}{{Tax and Duin}}{{}}}
\bibcite{bottleneck-nips}{{21}{2001}{{Tishby and Slonim}}{{}}}
\bibcite{Vapnik95}{{22}{1995}{{Vapnik}}{{}}}