| <!-- 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 1.8.14"/> |
| <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="navtreedata.js"></script> |
| <script type="text/javascript" src="navtree.js"></script> |
| <script type="text/javascript"> |
| /* @license magnet:?xt=urn:btih:cf05388f2679ee054f2beb29a391d25f4e673ac3&dn=gpl-2.0.txt GPL-v2 */ |
| $(document).ready(initResizable); |
| /* @license-end */</script> |
| <link href="search/search.css" rel="stylesheet" type="text/css"/> |
| <script type="text/javascript" src="search/searchdata.js"></script> |
| <script type="text/javascript" src="search/search.js"></script> |
| <script type="text/javascript"> |
| /* @license magnet:?xt=urn:btih:cf05388f2679ee054f2beb29a391d25f4e673ac3&dn=gpl-2.0.txt GPL-v2 */ |
| $(document).ready(function() { init_search(); }); |
| /* @license-end */ |
| </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 type="text/javascript" async src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/MathJax.js"></script> |
| <!-- hack in the navigation tree --> |
| <script type="text/javascript" src="eigen_navtree_hacks.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', 'madlib.apache.org'); |
| 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 id="projectlogo"><a href="http://madlib.apache.org"><img alt="Logo" src="madlib.png" height="50" style="padding-left:0.5em;" border="0"/ ></a></td> |
| <td style="padding-left: 0.5em;"> |
| <div id="projectname"> |
| <span id="projectnumber">1.15</span> |
| </div> |
| <div id="projectbrief">User Documentation for Apache MADlib</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 1.8.14 --> |
| <script type="text/javascript"> |
| /* @license magnet:?xt=urn:btih:cf05388f2679ee054f2beb29a391d25f4e673ac3&dn=gpl-2.0.txt GPL-v2 */ |
| var searchBox = new SearchBox("searchBox", "search",false,'Search'); |
| /* @license-end */ |
| </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"> |
| /* @license magnet:?xt=urn:btih:cf05388f2679ee054f2beb29a391d25f4e673ac3&dn=gpl-2.0.txt GPL-v2 */ |
| $(document).ready(function(){initNavTree('group__grp__kmeans.html','');}); |
| /* @license-end */ |
| </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)"> |
| </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__unsupervised.html">Unsupervised Learning</a> » <a class="el" href="group__grp__clustering.html">Clustering</a></div></div> </div> |
| </div><!--header--> |
| <div class="contents"> |
| <div class="toc"><b>Contents</b> <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="#assignment">Cluster Assignment</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, |
| seeding_sample_ratio |
| ) |
| </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 or an array expression.</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>seeding_sample_ratio (optional) </dt> |
| <dd><p class="startdd">DOUBLE PRECISION, default: 1.0. The proportion of subsample of original dataset to use for kmeans++ centroid seeding method. Kmeans++ scans through the data sequentially 'k' times and can be too slow for big datasets. When 'seeding_sample_ratio' is greater than 0 (thresholded to be maximum value of 1.0), the seeding is run on an uniform random subsample of the data. Note: the final K-means algorithm is run on the complete dataset. This parameter only builds a subsample for the seeding and is only available for kmeans++.</p> |
| <p class="enddd"></p> |
| </dd> |
| <dt>rel_initial_centroids </dt> |
| <dd><p class="startdd">TEXT. The set of initial centroids. </p> |
| <p class="enddd"></p> |
| </dd> |
| <dt>expr_centroid </dt> |
| <dd><p class="startdd">TEXT. The name of the column (or the array expression) 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>cluster_variance </th><td>DOUBLE PRECISION[]. The value of the objective function per cluster. </td></tr> |
| <tr> |
| <th>objective_fn </th><td>DOUBLE PRECISION. 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="assignment"></a></p><dl class="section user"><dt>Cluster Assignment</dt><dd></dd></dl> |
| <p>After training, the cluster assignment for each data point can be computed with the help of the following function:</p> |
| <pre class="syntax"> |
| closest_column( m, x ) |
| </pre><p><b>Argument</b> </p><dl class="arglist"> |
| <dt>m </dt> |
| <dd>DOUBLE PRECISION[][]. The learned centroids from the training function. </dd> |
| <dt>x </dt> |
| <dd>DOUBLE PRECISION[]. The data point. </dd> |
| </dl> |
| <p><b>Output format</b> </p><table class="output"> |
| <tr> |
| <th>column_id </th><td>INTEGER. The cluster assignment (zero-based). </td></tr> |
| <tr> |
| <th>distance </th><td>DOUBLE PRECISION. The distance to the cluster centroid. </td></tr> |
| </table> |
| <p><a class="anchor" id="examples"></a></p><dl class="section user"><dt>Examples</dt><dd></dd></dl> |
| <p>Note: Your results may not be exactly the same as below due to the nature of the k-means algorithm.</p> |
| <ol type="1"> |
| <li>Prepare some input data: <pre class="example"> |
| DROP TABLE IF EXISTS km_sample; |
| CREATE TABLE km_sample(pid int, points double precision[]); |
| INSERT INTO km_sample VALUES |
| (1, '{14.23, 1.71, 2.43, 15.6, 127, 2.8, 3.0600, 0.2800, 2.29, 5.64, 1.04, 3.92, 1065}'), |
| (2, '{13.2, 1.78, 2.14, 11.2, 1, 2.65, 2.76, 0.26, 1.28, 4.38, 1.05, 3.49, 1050}'), |
| (3, '{13.16, 2.36, 2.67, 18.6, 101, 2.8, 3.24, 0.3, 2.81, 5.6799, 1.03, 3.17, 1185}'), |
| (4, '{14.37, 1.95, 2.5, 16.8, 113, 3.85, 3.49, 0.24, 2.18, 7.8, 0.86, 3.45, 1480}'), |
| (5, '{13.24, 2.59, 2.87, 21, 118, 2.8, 2.69, 0.39, 1.82, 4.32, 1.04, 2.93, 735}'), |
| (6, '{14.2, 1.76, 2.45, 15.2, 112, 3.27, 3.39, 0.34, 1.97, 6.75, 1.05, 2.85, 1450}'), |
| (7, '{14.39, 1.87, 2.45, 14.6, 96, 2.5, 2.52, 0.3, 1.98, 5.25, 1.02, 3.58, 1290}'), |
| (8, '{14.06, 2.15, 2.61, 17.6, 121, 2.6, 2.51, 0.31, 1.25, 5.05, 1.06, 3.58, 1295}'), |
| (9, '{14.83, 1.64, 2.17, 14, 97, 2.8, 2.98, 0.29, 1.98, 5.2, 1.08, 2.85, 1045}'), |
| (10, '{13.86, 1.35, 2.27, 16, 98, 2.98, 3.15, 0.22, 1.8500, 7.2199, 1.01, 3.55, 1045}'); |
| </pre></li> |
| <li>Run k-means clustering using kmeans++ for centroid seeding: <pre class="example"> |
| DROP TABLE IF EXISTS km_result; |
| -- Run kmeans algorithm |
| CREATE TABLE km_result AS |
| SELECT * FROM madlib.kmeanspp('km_sample', 'points', 2, |
| 'madlib.squared_dist_norm2', |
| 'madlib.avg', 20, 0.001); |
| \x on; |
| SELECT * FROM km_result; |
| </pre> Result: <pre class="result"> |
| centroids | {{14.036,2.018,2.536,16.56,108.6,3.004,3.03,0.298,2.038,6.10598,1.004,3.326,1340},{13.872,1.814,2.376,15.56,88.2,2.806,2.928,0.288,1.844,5.35198,1.044,3.348,988}} |
| cluster_variance | {60672.638245208,90512.324426408} |
| objective_fn | 151184.962671616 |
| frac_reassigned | 0 |
| num_iterations | 2 |
| </pre></li> |
| <li>Calculate the simplified silhouette coefficient: <pre class="example"> |
| SELECT * FROM madlib.simple_silhouette( 'km_sample', |
| 'points', |
| (SELECT centroids FROM km_result), |
| 'madlib.dist_norm2' |
| ); |
| </pre> Result: <pre class="result"> |
| simple_silhouette | 0.68978804882941 |
| </pre></li> |
| <li>Find the cluster assignment for each point: <pre class="example"> |
| \x off; |
| -- Get point assignment |
| SELECT data.*, (madlib.closest_column(centroids, points)).column_id as cluster_id |
| FROM km_sample as data, km_result |
| ORDER BY data.pid; |
| </pre> Result: <pre class="result"> |
| pid | points | cluster_id |
| -----+--------------------------------------------------------------------+------------ |
| 1 | {14.23,1.71,2.43,15.6,127,2.8,3.06,0.28,2.29,5.64,1.04,3.92,1065} | 1 |
| 2 | {13.2,1.78,2.14,11.2,1,2.65,2.76,0.26,1.28,4.38,1.05,3.49,1050} | 1 |
| 3 | {13.16,2.36,2.67,18.6,101,2.8,3.24,0.3,2.81,5.6799,1.03,3.17,1185} | 0 |
| 4 | {14.37,1.95,2.5,16.8,113,3.85,3.49,0.24,2.18,7.8,0.86,3.45,1480} | 0 |
| 5 | {13.24,2.59,2.87,21,118,2.8,2.69,0.39,1.82,4.32,1.04,2.93,735} | 1 |
| 6 | {14.2,1.76,2.45,15.2,112,3.27,3.39,0.34,1.97,6.75,1.05,2.85,1450} | 0 |
| 7 | {14.39,1.87,2.45,14.6,96,2.5,2.52,0.3,1.98,5.25,1.02,3.58,1290} | 0 |
| 8 | {14.06,2.15,2.61,17.6,121,2.6,2.51,0.31,1.25,5.05,1.06,3.58,1295} | 0 |
| 9 | {14.83,1.64,2.17,14,97,2.8,2.98,0.29,1.98,5.2,1.08,2.85,1045} | 1 |
| 10 | {13.86,1.35,2.27,16,98,2.98,3.15,0.22,1.85,7.2199,1.01,3.55,1045} | 1 |
| (10 rows) |
| </pre></li> |
| <li>Unnest the cluster centroids 2-D array to get a set of 1-D centroid arrays: <pre class="example"> |
| DROP TABLE IF EXISTS km_centroids_unnest; |
| -- Run unnest function |
| CREATE TABLE km_centroids_unnest AS |
| SELECT (madlib.array_unnest_2d_to_1d(centroids)).* |
| FROM km_result; |
| SELECT * FROM km_centroids_unnest ORDER BY 1; |
| </pre> Result: <pre class="result"> |
| unnest_row_id | unnest_result |
| ---------------+---------------------------------------------------------------------------------- |
| 1 | {14.036,2.018,2.536,16.56,108.6,3.004,3.03,0.298,2.038,6.10598,1.004,3.326,1340} |
| 2 | {13.872,1.814,2.376,15.56,88.2,2.806,2.928,0.288,1.844,5.35198,1.044,3.348,988} |
| (2 rows) |
| </pre> Note that the ID column returned by <a class="el" href="array__ops_8sql__in.html#af057b589f2a2cb1095caa99feaeb3d70" title="This function takes a 2-D array as the input and unnests it by one level. It returns a set of 1-D arr...">array_unnest_2d_to_1d()</a> is not guaranteed to be the same as the cluster ID assigned by k-means. See below to create the correct cluster IDs.</li> |
| <li>Create cluster IDs for 1-D centroid arrays so that cluster ID for any centroid can be matched to the cluster assignment for the data points: <pre class="example"> |
| SELECT cent.*, (madlib.closest_column(centroids, unnest_result)).column_id as cluster_id |
| FROM km_centroids_unnest as cent, km_result |
| ORDER BY cent.unnest_row_id; |
| </pre> Result: <pre class="result"> |
| unnest_row_id | unnest_result | cluster_id |
| ---------------+----------------------------------------------------------------------------------+------------ |
| 1 | {14.036,2.018,2.536,16.56,108.6,3.004,3.03,0.298,2.038,6.10598,1.004,3.326,1340} | 0 |
| 2 | {13.872,1.814,2.376,15.56,88.2,2.806,2.928,0.288,1.844,5.35198,1.044,3.348,988} | 1 |
| (2 rows) |
| </pre></li> |
| <li>Run the same example as above, but using array input. Create the input table: <pre class="example"> |
| DROP TABLE IF EXISTS km_arrayin CASCADE; |
| CREATE TABLE km_arrayin(pid int, |
| p1 float, |
| p2 float, |
| p3 float, |
| p4 float, |
| p5 float, |
| p6 float, |
| p7 float, |
| p8 float, |
| p9 float, |
| p10 float, |
| p11 float, |
| p12 float, |
| p13 float); |
| INSERT INTO km_arrayin VALUES |
| (1, 14.23, 1.71, 2.43, 15.6, 127, 2.8, 3.0600, 0.2800, 2.29, 5.64, 1.04, 3.92, 1065), |
| (2, 13.2, 1.78, 2.14, 11.2, 1, 2.65, 2.76, 0.26, 1.28, 4.38, 1.05, 3.49, 1050), |
| (3, 13.16, 2.36, 2.67, 18.6, 101, 2.8, 3.24, 0.3, 2.81, 5.6799, 1.03, 3.17, 1185), |
| (4, 14.37, 1.95, 2.5, 16.8, 113, 3.85, 3.49, 0.24, 2.18, 7.8, 0.86, 3.45, 1480), |
| (5, 13.24, 2.59, 2.87, 21, 118, 2.8, 2.69, 0.39, 1.82, 4.32, 1.04, 2.93, 735), |
| (6, 14.2, 1.76, 2.45, 15.2, 112, 3.27, 3.39, 0.34, 1.97, 6.75, 1.05, 2.85, 1450), |
| (7, 14.39, 1.87, 2.45, 14.6, 96, 2.5, 2.52, 0.3, 1.98, 5.25, 1.02, 3.58, 1290), |
| (8, 14.06, 2.15, 2.61, 17.6, 121, 2.6, 2.51, 0.31, 1.25, 5.05, 1.06, 3.58, 1295), |
| (9, 14.83, 1.64, 2.17, 14, 97, 2.8, 2.98, 0.29, 1.98, 5.2, 1.08, 2.85, 1045), |
| (10, 13.86, 1.35, 2.27, 16, 98, 2.98, 3.15, 0.22, 1.8500, 7.2199, 1.01, 3.55, 1045); |
| </pre> Now find the cluster assignment for each point: <pre class="example"> |
| DROP TABLE IF EXISTS km_result; |
| -- Run kmeans algorithm |
| CREATE TABLE km_result AS |
| SELECT * FROM madlib.kmeans_random('km_arrayin', |
| 'ARRAY[p1, p2, p3, p4, p5, p6, |
| p7, p8, p9, p10, p11, p12, p13]', |
| 2, |
| 'madlib.squared_dist_norm2', |
| 'madlib.avg', |
| 20, |
| 0.001); |
| -- Get point assignment |
| SELECT data.*, (madlib.closest_column(centroids, |
| ARRAY[p1, p2, p3, p4, p5, p6, |
| p7, p8, p9, p10, p11, p12, p13])).column_id as cluster_id |
| FROM km_arrayin as data, km_result |
| ORDER BY data.pid; |
| </pre> This produces the result in column format: <pre class="result"> |
| pid | p1 | p2 | p3 | p4 | p5 | p6 | p7 | p8 | p9 | p10 | p11 | p12 | p13 | cluster_id |
| -----+-------+------+------+------+-----+------+------+------+------+--------+------+------+------+------------ |
| 1 | 14.23 | 1.71 | 2.43 | 15.6 | 127 | 2.8 | 3.06 | 0.28 | 2.29 | 5.64 | 1.04 | 3.92 | 1065 | 0 |
| 2 | 13.2 | 1.78 | 2.14 | 11.2 | 1 | 2.65 | 2.76 | 0.26 | 1.28 | 4.38 | 1.05 | 3.49 | 1050 | 0 |
| 3 | 13.16 | 2.36 | 2.67 | 18.6 | 101 | 2.8 | 3.24 | 0.3 | 2.81 | 5.6799 | 1.03 | 3.17 | 1185 | 0 |
| 4 | 14.37 | 1.95 | 2.5 | 16.8 | 113 | 3.85 | 3.49 | 0.24 | 2.18 | 7.8 | 0.86 | 3.45 | 1480 | 1 |
| 5 | 13.24 | 2.59 | 2.87 | 21 | 118 | 2.8 | 2.69 | 0.39 | 1.82 | 4.32 | 1.04 | 2.93 | 735 | 0 |
| 6 | 14.2 | 1.76 | 2.45 | 15.2 | 112 | 3.27 | 3.39 | 0.34 | 1.97 | 6.75 | 1.05 | 2.85 | 1450 | 1 |
| 7 | 14.39 | 1.87 | 2.45 | 14.6 | 96 | 2.5 | 2.52 | 0.3 | 1.98 | 5.25 | 1.02 | 3.58 | 1290 | 1 |
| 8 | 14.06 | 2.15 | 2.61 | 17.6 | 121 | 2.6 | 2.51 | 0.31 | 1.25 | 5.05 | 1.06 | 3.58 | 1295 | 1 |
| 9 | 14.83 | 1.64 | 2.17 | 14 | 97 | 2.8 | 2.98 | 0.29 | 1.98 | 5.2 | 1.08 | 2.85 | 1045 | 0 |
| 10 | 13.86 | 1.35 | 2.27 | 16 | 98 | 2.98 | 3.15 | 0.22 | 1.85 | 7.2199 | 1.01 | 3.55 | 1045 | 0 |
| (10 rows) |
| </pre></li> |
| </ol> |
| <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>simple_silhouette()</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 Mon Aug 6 2018 21:55:39 for MADlib by |
| <a href="http://www.doxygen.org/index.html"> |
| <img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.14 </li> |
| </ul> |
| </div> |
| </body> |
| </html> |
