Distributed K-Means Clustering

K-Means clustering is the popular and easy to understand clustering algorithm. Both Spark MLlib and Apache Mahout have implemented the algorithm. However, Mahout MapReduce version of K-Means algorithm is slow compared to Spark. This is due to the huge IO involved in every iteration of the algorithm. Basically, at each iteration, the intermediate output results (clusters)  have to be written to the disk and then read in again together with input dataset from the disk. The good news is Apache Mahout has started the work to integrate with Spark.

Let’s look at the codes. KMeansDriver is the driver class. User can choose to run a MapReduce or sequential version of the algorithm. The program arguments are the path to the initial cluster centers, dataset input path,  convergence rate, distance measure implementation class, max number of iterations, and etc.

  public int run(String[] args) throws Exception {
            "The input centroids, as Vectors.  Must be a SequenceFile of Writable, Cluster/Canopy.  "
                + "If k is also specified, then a random set of vectors will be selected"
                + " and written out to this path first").create());
            "The k in k-Means.  If specified, then a random selection of k Vectors will be chosen"
                + " as the Centroid and written to the clusters input path.").create());
    if (parseArguments(args) == null) {
      return -1;
    Path input = getInputPath();
    Path clusters = new Path(getOption(DefaultOptionCreator.CLUSTERS_IN_OPTION));
    Path output = getOutputPath();
    String measureClass = getOption(DefaultOptionCreator.DISTANCE_MEASURE_OPTION);
    if (measureClass == null) {
      measureClass = SquaredEuclideanDistanceMeasure.class.getName();
    double convergenceDelta = Double.parseDouble(getOption(DefaultOptionCreator.CONVERGENCE_DELTA_OPTION));
    int maxIterations = Integer.parseInt(getOption(DefaultOptionCreator.MAX_ITERATIONS_OPTION));
    if (hasOption(DefaultOptionCreator.OVERWRITE_OPTION)) {
      HadoopUtil.delete(getConf(), output);
    DistanceMeasure measure = ClassUtils.instantiateAs(measureClass, DistanceMeasure.class);
    if (hasOption(DefaultOptionCreator.NUM_CLUSTERS_OPTION)) {
      int numClusters = Integer.parseInt(getOption(DefaultOptionCreator.NUM_CLUSTERS_OPTION));

      Long seed = null;
      if (hasOption(DefaultOptionCreator.RANDOM_SEED)) {
        seed = Long.parseLong(getOption(DefaultOptionCreator.RANDOM_SEED));

      clusters = RandomSeedGenerator.buildRandom(getConf(), input, clusters, numClusters, measure, seed);
    boolean runClustering = hasOption(DefaultOptionCreator.CLUSTERING_OPTION);
    boolean runSequential = getOption(DefaultOptionCreator.METHOD_OPTION).equalsIgnoreCase(
    double clusterClassificationThreshold = 0.0;
    if (hasOption(DefaultOptionCreator.OUTLIER_THRESHOLD)) {
      clusterClassificationThreshold = Double.parseDouble(getOption(DefaultOptionCreator.OUTLIER_THRESHOLD));
    run(getConf(), input, clusters, output, convergenceDelta, maxIterations, runClustering,
        clusterClassificationThreshold, runSequential);
    return 0;

The main MapReduce clustering logic can be found below. As you seen, it is an iterative process where in each iteration, we create a MapReduce job that reads in the input vectors and previous clusters and start assign the input vectors to the closest cluster until either we reach convergence or hit the specified number of iterations. There is a lot of IO overhead here. Each iteration, the MapReduce job has to read the previous results which are the clusters and input vectors all over again. (Take a look at Kluster.java which represents a cluster with cluster id, the center and the associated distance measure)

   * Iterate over data using a prior-trained ClusterClassifier, for a number of iterations using a mapreduce
   * implementation
   * @param conf
   *          the Configuration
   * @param inPath
   *          a Path to input VectorWritables
   * @param priorPath
   *          a Path to the prior classifier
   * @param outPath
   *          a Path of output directory
   * @param numIterations
   *          the int number of iterations to perform
  public static void iterateMR(Configuration conf, Path inPath, Path priorPath, Path outPath, int numIterations)
    throws IOException, InterruptedException, ClassNotFoundException {
    ClusteringPolicy policy = ClusterClassifier.readPolicy(priorPath);
    Path clustersOut = null;
    int iteration = 1;
    while (iteration <= numIterations) {
      conf.set(PRIOR_PATH_KEY, priorPath.toString());
      String jobName = "Cluster Iterator running iteration " + iteration + " over priorPath: " + priorPath;
      Job job = new Job(conf, jobName);
      FileInputFormat.addInputPath(job, inPath);
      clustersOut = new Path(outPath, Cluster.CLUSTERS_DIR + iteration);
      priorPath = clustersOut;
      FileOutputFormat.setOutputPath(job, clustersOut);
      if (!job.waitForCompletion(true)) {
        throw new InterruptedException("Cluster Iteration " + iteration + " failed processing " + priorPath);
      ClusterClassifier.writePolicy(policy, clustersOut);
      FileSystem fs = FileSystem.get(outPath.toUri(), conf);
      if (isConverged(clustersOut, conf, fs)) {
    Path finalClustersIn = new Path(outPath, Cluster.CLUSTERS_DIR + (iteration - 1) + Cluster.FINAL_ITERATION_SUFFIX);
    FileSystem.get(clustersOut.toUri(), conf).rename(clustersOut, finalClustersIn);

Stay tuned for more infos about the Mapper and Reducer implementation of the algorithm.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s