| <!-- HTML header for doxygen 1.8.4--> |
| <!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd"> |
| <html xmlns="http://www.w3.org/1999/xhtml"> |
| <head> |
| <meta http-equiv="Content-Type" content="text/xhtml;charset=UTF-8"/> |
| <meta http-equiv="X-UA-Compatible" content="IE=9"/> |
| <meta name="generator" content="Doxygen 1.8.4"/> |
| <meta name="keywords" content="madlib,postgres,greenplum,machine learning,data mining,deep learning,ensemble methods,data science,market basket analysis,affinity analysis,pca,lda,regression,elastic net,huber white,proportional hazards,k-means,latent dirichlet allocation,bayes,support vector machines,svm"/> |
| <title>MADlib: k-Means Clustering</title> |
| <link href="tabs.css" rel="stylesheet" type="text/css"/> |
| <script type="text/javascript" src="jquery.js"></script> |
| <script type="text/javascript" src="dynsections.js"></script> |
| <link href="navtree.css" rel="stylesheet" type="text/css"/> |
| <script type="text/javascript" src="resize.js"></script> |
| <script type="text/javascript" src="navtree.js"></script> |
| <script type="text/javascript"> |
| $(document).ready(initResizable); |
| $(window).load(resizeHeight); |
| </script> |
| <link href="search/search.css" rel="stylesheet" type="text/css"/> |
| <script type="text/javascript" src="search/search.js"></script> |
| <script type="text/javascript"> |
| $(document).ready(function() { searchBox.OnSelectItem(0); }); |
| </script> |
| <script type="text/x-mathjax-config"> |
| MathJax.Hub.Config({ |
| extensions: ["tex2jax.js", "TeX/AMSmath.js", "TeX/AMSsymbols.js"], |
| jax: ["input/TeX","output/HTML-CSS"], |
| }); |
| </script><script src="../mathjax/MathJax.js"></script> |
| <link href="doxygen.css" rel="stylesheet" type="text/css" /> |
| <link href="madlib_extra.css" rel="stylesheet" type="text/css"/> |
| <!-- google analytics --> |
| <script> |
| (function(i,s,o,g,r,a,m){i['GoogleAnalyticsObject']=r;i[r]=i[r]||function(){ |
| (i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o), |
| m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;m.parentNode.insertBefore(a,m) |
| })(window,document,'script','//www.google-analytics.com/analytics.js','ga'); |
| ga('create', 'UA-45382226-1', 'auto'); |
| ga('send', 'pageview'); |
| </script> |
| </head> |
| <body> |
| <div id="top"><!-- do not remove this div, it is closed by doxygen! --> |
| <div id="titlearea"> |
| <table cellspacing="0" cellpadding="0"> |
| <tbody> |
| <tr style="height: 56px;"> |
| <td style="padding-left: 0.5em;"> |
| <div id="projectname">MADlib |
|  <span id="projectnumber">1.4.1</span> |
| </div> |
| <div id="projectbrief">User Documentation</div> |
| </td> |
| <td> <div id="MSearchBox" class="MSearchBoxInactive"> |
| <span class="left"> |
| <img id="MSearchSelect" src="search/mag_sel.png" |
| onmouseover="return searchBox.OnSearchSelectShow()" |
| onmouseout="return searchBox.OnSearchSelectHide()" |
| alt=""/> |
| <input type="text" id="MSearchField" value="Search" accesskey="S" |
| onfocus="searchBox.OnSearchFieldFocus(true)" |
| onblur="searchBox.OnSearchFieldFocus(false)" |
| onkeyup="searchBox.OnSearchFieldChange(event)"/> |
| </span><span class="right"> |
| <a id="MSearchClose" href="javascript:searchBox.CloseResultsWindow()"><img id="MSearchCloseImg" border="0" src="search/close.png" alt=""/></a> |
| </span> |
| </div> |
| </td> |
| </tr> |
| </tbody> |
| </table> |
| </div> |
| <!-- end header part --> |
| <!-- Generated by Doxygen 1.8.4 --> |
| <script type="text/javascript"> |
| var searchBox = new SearchBox("searchBox", "search",false,'Search'); |
| </script> |
| </div><!-- top --> |
| <div id="side-nav" class="ui-resizable side-nav-resizable"> |
| <div id="nav-tree"> |
| <div id="nav-tree-contents"> |
| <div id="nav-sync" class="sync"></div> |
| </div> |
| </div> |
| <div id="splitbar" style="-moz-user-select:none;" |
| class="ui-resizable-handle"> |
| </div> |
| </div> |
| <script type="text/javascript"> |
| $(document).ready(function(){initNavTree('group__grp__kmeans.html','');}); |
| </script> |
| <div id="doc-content"> |
| <!-- window showing the filter options --> |
| <div id="MSearchSelectWindow" |
| onmouseover="return searchBox.OnSearchSelectShow()" |
| onmouseout="return searchBox.OnSearchSelectHide()" |
| onkeydown="return searchBox.OnSearchSelectKey(event)"> |
| <a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(0)"><span class="SelectionMark"> </span>All</a><a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(1)"><span class="SelectionMark"> </span>Files</a><a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(2)"><span class="SelectionMark"> </span>Functions</a><a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(3)"><span class="SelectionMark"> </span>Variables</a><a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(4)"><span class="SelectionMark"> </span>Groups</a></div> |
| |
| <!-- iframe showing the search results (closed by default) --> |
| <div id="MSearchResultsWindow"> |
| <iframe src="javascript:void(0)" frameborder="0" |
| name="MSearchResults" id="MSearchResults"> |
| </iframe> |
| </div> |
| |
| <div class="header"> |
| <div class="headertitle"> |
| <div class="title">k-Means Clustering<div class="ingroups"><a class="el" href="group__grp__clustering.html">Clustering</a></div></div> </div> |
| </div><!--header--> |
| <div class="contents"> |
| <div class="toc"><b>Contents</b> </p> |
| <ul> |
| <li class="level1"> |
| <a href="#train">Training Function</a> </li> |
| <li class="level1"> |
| <a href="#output">Output Format</a> </li> |
| <li class="level1"> |
| <a href="#examples">Examples</a> </li> |
| <li class="level1"> |
| <a href="#notes">Notes</a> </li> |
| <li class="level1"> |
| <a href="#background">Technical Background</a> </li> |
| <li class="level1"> |
| <a href="#literature">Literature</a> </li> |
| <li class="level1"> |
| <a href="#related">Related Topics</a> </li> |
| </ul> |
| </div><p>Clustering refers to the problem of partitioning a set of objects according to some problem-dependent measure of <em>similarity</em>. In the k-means variant, given \( n \) points \( x_1, \dots, x_n \in \mathbb R^d \), the goal is to position \( k \) centroids \( c_1, \dots, c_k \in \mathbb R^d \) so that the sum of <em>distances</em> between each point and its closest centroid is minimized. Each centroid represents a cluster that consists of all points to which this centroid is closest.</p> |
| <p><a class="anchor" id="train"></a></p> |
| <dl class="section user"><dt>Training Function</dt><dd></dd></dl> |
| <p>The k-means algorithm can be invoked in four ways, depending on the source of the initial set of centroids:</p> |
| <ul> |
| <li>Use the random centroid seeding method. <pre class="syntax"> |
| kmeans_random( rel_source, |
| expr_point, |
| k, |
| fn_dist, |
| agg_centroid, |
| max_num_iterations, |
| min_frac_reassigned |
| ) |
| </pre></li> |
| <li>Use the kmeans++ centroid seeding method. <pre class="syntax"> |
| kmeanspp( rel_source, |
| expr_point, |
| k, |
| fn_dist, |
| agg_centroid, |
| max_num_iterations, |
| min_frac_reassigned |
| ) |
| </pre></li> |
| <li>Supply an initial centroid set in a relation identified by the <em>rel_initial_centroids</em> argument. <pre class="syntax"> |
| kmeans( rel_source, |
| expr_point, |
| rel_initial_centroids, |
| expr_centroid, |
| fn_dist, |
| agg_centroid, |
| max_num_iterations, |
| min_frac_reassigned |
| ) |
| </pre></li> |
| <li>Provide an initial centroid set as an array expression in the <em>initial_centroids</em> argument. <pre class="syntax"> |
| kmeans( rel_source, |
| expr_point, |
| initial_centroids, |
| fn_dist, |
| agg_centroid, |
| max_num_iterations, |
| min_frac_reassigned |
| ) |
| </pre> <b>Arguments</b> <dl class="arglist"> |
| <dt>rel_source </dt> |
| <dd><p class="startdd">TEXT. The name of the table containing the input data points.</p> |
| <p>Data points and predefined centroids (if used) are expected to be stored row-wise, in a column of type <code><a class="el" href="group__grp__svec.html">SVEC</a></code> (or any type convertible to <code><a class="el" href="group__grp__svec.html">SVEC</a></code>, like <code>FLOAT[]</code> or <code>INTEGER[]</code>). Data points with non-finite values (NULL, NaN, infinity) in any component are skipped during analysis. </p> |
| <p class="enddd"></p> |
| </dd> |
| <dt>expr_point </dt> |
| <dd><p class="startdd">TEXT. The name of the column with point coordinates.</p> |
| <p class="enddd"></p> |
| </dd> |
| <dt>k </dt> |
| <dd><p class="startdd">INTEGER. The number of centroids to calculate.</p> |
| <p class="enddd"></p> |
| </dd> |
| <dt>fn_dist (optional) </dt> |
| <dd><p class="startdd">TEXT, default: squared_dist_norm2'. The name of the function to use to calculate the distance from a data point to a centroid.</p> |
| <p>The following distance functions can be used (computation of barycenter/mean in parentheses): </p> |
| <ul> |
| <li> |
| <b><a class="el" href="linalg_8sql__in.html#aad193850e79c4b9d811ca9bc53e13476">dist_norm1</a></b>: 1-norm/Manhattan (element-wise median [Note that MADlib does not provide a median aggregate function for support and performance reasons.]) </li> |
| <li> |
| <b><a class="el" href="linalg_8sql__in.html#aa58e51526edea6ea98db30b6f250adb4">dist_norm2</a></b>: 2-norm/Euclidean (element-wise mean) </li> |
| <li> |
| <b><a class="el" href="linalg_8sql__in.html#a00a08e69f27524f2096032214e15b668">squared_dist_norm2</a></b>: squared Euclidean distance (element-wise mean) </li> |
| <li> |
| <b><a class="el" href="linalg_8sql__in.html#a8c7b9281a72ff22caf06161701b27e84">dist_angle</a></b>: angle (element-wise mean of normalized points) </li> |
| <li> |
| <b><a class="el" href="linalg_8sql__in.html#afa13b4c6122b99422d666dedea136c18">dist_tanimoto</a></b>: tanimoto (element-wise mean of normalized points <a href="#kmeans-lit-5">[5]</a>) </li> |
| <li> |
| <b>user defined function</b> with signature <code>DOUBLE PRECISION[] x, DOUBLE PRECISION[] y -> DOUBLE PRECISION</code></li> |
| </ul> |
| <p class="enddd"></p> |
| </dd> |
| <dt>agg_centroid (optional) </dt> |
| <dd><p class="startdd">TEXT, default: 'avg'. The name of the aggregate function used to determine centroids.</p> |
| <p>The following aggregate functions can be used:</p> |
| <ul> |
| <li> |
| <b><a class="el" href="linalg_8sql__in.html#a1aa37f73fb1cd8d7d106aa518dd8c0b4">avg</a></b>: average (Default) </li> |
| <li> |
| <b><a class="el" href="linalg_8sql__in.html#a0b04663ca206f03e66aed5ea2b4cc461">normalized_avg</a></b>: normalized average</li> |
| </ul> |
| <p class="enddd"></p> |
| </dd> |
| <dt>max_num_iterations (optional) </dt> |
| <dd><p class="startdd">INTEGER, default: 20. The maximum number of iterations to perform.</p> |
| <p class="enddd"></p> |
| </dd> |
| <dt>min_frac_reassigned (optional) </dt> |
| <dd><p class="startdd">DOUBLE PRECISION, default: 0.001. The minimum fraction of centroids reassigned to continue iterating. When fewer than this fraction of centroids are reassigned in an iteration, the calculation completes.</p> |
| <p class="enddd"></p> |
| </dd> |
| <dt>rel_initial_centroids </dt> |
| <dd><p class="startdd">TEXT. The set of initial centroids. The centroid relation is expected to be of the following form: </p> |
| <pre> |
| {TABLE|VIEW} rel_initial_centroids ( |
| ... |
| expr_centroid DOUBLE PRECISION[], |
| ... |
| ) |
| </pre><p> where <em>expr_centroid</em> is the name of a column with coordinates. </p> |
| <p class="enddd"></p> |
| </dd> |
| <dt>expr_centroid </dt> |
| <dd><p class="startdd">TEXT. The name of the column in the <em>rel_initial_centroids</em> relation that contains the centroid coordinates.</p> |
| <p class="enddd"></p> |
| </dd> |
| <dt>initial_centroids </dt> |
| <dd>TEXT. A string containing a DOUBLE PRECISION array expression with the initial centroid coordinates. </dd> |
| </dl> |
| </li> |
| </ul> |
| <p><a class="anchor" id="output"></a></p> |
| <dl class="section user"><dt>Output Format</dt><dd></dd></dl> |
| <p>The output of the k-means module is a composite type with the following columns: </p> |
| <table class="output"> |
| <tr> |
| <th>centroids </th><td>DOUBLE PRECISION[][]. The final centroid positions. </td></tr> |
| <tr> |
| <th>objective_fn </th><td>DOUBLE PRECISON. The value of the objective function. </td></tr> |
| <tr> |
| <th>frac_reassigned </th><td>DOUBLE PRECISION. The fraction of points reassigned in the last iteration. </td></tr> |
| <tr> |
| <th>num_iterations </th><td>INTEGER. The total number of iterations executed. </td></tr> |
| </table> |
| <p><a class="anchor" id="examples"></a></p> |
| <dl class="section user"><dt>Examples</dt><dd><ol type="1"> |
| <li>Prepare some input data. <pre class="example"> |
| SELECT * FROM public.km_sample LIMIT 5; |
| </pre> Result: <pre class="result"> |
| points |
|  ------------------------------------------ |
| {1,1}:{15.8822241332382,105.945462542586} |
| {1,1}:{34.5065216883086,72.3126099305227} |
| {1,1}:{22.5074400822632,95.3209559689276} |
| {1,1}:{70.2589857042767,68.7395178806037} |
| {1,1}:{30.9844257542863,25.3213323024102} |
| (5 rows) |
| </pre> Note: the example <em>points</em> is type <code><a class="el" href="group__grp__svec.html">SVEC</a></code>.</li> |
| <li>Run k-means clustering using kmeans++ for centroid seeding: <pre class="example"> |
| SELECT * FROM madlib.kmeanspp( 'km_sample', |
| 'points', |
| 2, |
| 'madlib.squared_dist_norm2', |
| 'madlib.avg', |
| 20, |
| 0.001 |
| ); |
| </pre> Result: <pre class="result"> |
| centroids | objective_fn | frac_reassigned | num_iterations |
|  ------------------------------------------------------------------------+------------------+-----------------+---------------- |
| {{68.01668579784,48.9667382972952},{28.1452167573446,84.5992507653263}} | 586729.010675982 | 0.001 | 5 |
| </pre></li> |
| <li>Calculate the simplified silhouette coefficient: <pre class="example"> |
| SELECT * FROM madlib.simple_silhouette( 'km_test_svec', |
| 'points', |
| (SELECT centroids FROM |
| madlib.kmeanspp( 'km_test_svec', |
| 'points', |
| 2, |
| 'madlib.squared_dist_norm2', |
| 'madlib.avg', |
| 20, |
| 0.001)), |
| 'madlib.dist_norm2' |
| ); |
| </pre> Result: <pre class="result"> |
| simple_silhouette |
|  ------------------ |
| 0.611022970398174 |
| </pre></li> |
| </ol> |
| </dd></dl> |
| <p><a class="anchor" id="notes"></a></p> |
| <dl class="section user"><dt>Notes</dt><dd></dd></dl> |
| <p>The algorithm stops when one of the following conditions is met:</p> |
| <ul> |
| <li>The fraction of updated points is smaller than the convergence threshold (<em>min_frac_reassigned</em> argument). (Default: 0.001).</li> |
| <li>The algorithm reaches the maximum number of allowed iterations (<em>max_num_iterations</em> argument). (Default: 20).</li> |
| </ul> |
| <p>A popular method to assess the quality of the clustering is the <em>silhouette coefficient</em>, a simplified version of which is provided as part of the k-means module. Note that for large data sets, this computation is expensive.</p> |
| <p>The silhouette function has the following syntax: </p> |
| <pre class="syntax"> |
| simple_silhouette( rel_source, |
| expr_point, |
| centroids, |
| fn_dist |
| ) |
| </pre><p> <b>Arguments</b> </p> |
| <dl class="arglist"> |
| <dt>rel_source </dt> |
| <dd>TEXT. The name of the relation containing the input point. </dd> |
| <dt>expr_point </dt> |
| <dd>TEXT. An expression evaluating to point coordinates for each row in the relation. </dd> |
| <dt>centroids </dt> |
| <dd>TEXT. An expression evaluating to an array of centroids. </dd> |
| <dt>fn_dist (optional) </dt> |
| <dd>TEXT, default 'dist_norm2', The name of a function to calculate the distance of a point from a centroid. See the <em>fn_dist</em> argument of the k-means training function. </dd> |
| </dl> |
| <p><a class="anchor" id="background"></a></p> |
| <dl class="section user"><dt>Technical Background</dt><dd></dd></dl> |
| <p>Formally, we wish to minimize the following objective function: </p> |
| <p class="formulaDsp"> |
| \[ (c_1, \dots, c_k) \mapsto \sum_{i=1}^n \min_{j=1}^k \operatorname{dist}(x_i, c_j) \] |
| </p> |
| <p> In the most common case, \( \operatorname{dist} \) is the square of the Euclidean distance.</p> |
| <p>This problem is computationally difficult (NP-hard), yet the local-search heuristic proposed by Lloyd [4] performs reasonably well in practice. In fact, it is so ubiquitous today that it is often referred to as the <em>standard algorithm</em> or even just the <em>k-means algorithm</em> [1]. It works as follows:</p> |
| <ol type="1"> |
| <li>Seed the \( k \) centroids (see below)</li> |
| <li>Repeat until convergence:<ol type="a"> |
| <li>Assign each point to its closest centroid</li> |
| <li>Move each centroid to a position that minimizes the sum of distances in this cluster</li> |
| </ol> |
| </li> |
| <li>Convergence is achieved when no points change their assignments during step 2a.</li> |
| </ol> |
| <p>Since the objective function decreases in every step, this algorithm is guaranteed to converge to a local optimum.</p> |
| <p><a class="anchor" id="literature"></a></p> |
| <dl class="section user"><dt>Literature</dt><dd></dd></dl> |
| <p><a class="anchor" id="kmeans-lit-1"></a>[1] Wikipedia, K-means Clustering, <a href="http://en.wikipedia.org/wiki/K-means_clustering">http://en.wikipedia.org/wiki/K-means_clustering</a></p> |
| <p><a class="anchor" id="kmeans-lit-2"></a>[2] David Arthur, Sergei Vassilvitskii: k-means++: the advantages of careful seeding, Proceedings of the 18th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA'07), pp. 1027-1035, <a href="http://www.stanford.edu/~darthur/kMeansPlusPlus.pdf">http://www.stanford.edu/~darthur/kMeansPlusPlus.pdf</a></p> |
| <p><a class="anchor" id="kmeans-lit-3"></a>[3] E. R. Hruschka, L. N. C. Silva, R. J. G. B. Campello: Clustering Gene-Expression Data: A Hybrid Approach that Iterates Between k-Means and Evolutionary Search. In: Studies in Computational Intelligence - Hybrid Evolutionary Algorithms. pp. 313-335. Springer. 2007.</p> |
| <p><a class="anchor" id="kmeans-lit-4"></a>[4] Lloyd, Stuart: Least squares quantization in PCM. Technical Note, Bell Laboratories. Published much later in: IEEE Transactions on Information Theory 28(2), pp. 128-137. 1982.</p> |
| <p><a class="anchor" id="kmeans-lit-5"></a>[5] Leisch, Friedrich: A Toolbox for K-Centroids Cluster Analysis. In: Computational Statistics and Data Analysis, 51(2). pp. 526-544. 2006.</p> |
| <p><a class="anchor" id="related"></a></p> |
| <dl class="section user"><dt>Related Topics</dt><dd></dd></dl> |
| <p>File <a class="el" href="kmeans_8sql__in.html" title="Set of functions for k-means clustering. ">kmeans.sql_in</a> documenting the k-Means SQL functions</p> |
| <p><a class="el" href="group__grp__svec.html">Sparse Vectors</a></p> |
| <p><a class="el" href="kmeans_8sql__in.html#a71e7675758c99acbe7785819b6a85a8f">simple_silhouette()</a> </p> |
| </div><!-- contents --> |
| </div><!-- doc-content --> |
| <!-- start footer part --> |
| <div id="nav-path" class="navpath"><!-- id is needed for treeview function! --> |
| <ul> |
| <li class="footer">Generated on Thu Jan 9 2014 20:25:07 for MADlib by |
| <a href="http://www.doxygen.org/index.html"> |
| <img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.4 </li> |
| </ul> |
| </div> |
| </body> |
| </html> |