DMelt:AI/Data Clustering
Data clustering
Cluster analysis is the task of grouping a set of objects in such a way that objects in the same group (called a cluster) are more similar to each other than to those in other groups. DataMelt contains a framework for clustering analysis, i.e. for non-supervised learning in which the classification process does not depend on a priory information. It includes the following algorithms:
- K-means clustering analysis (single and multi pass)
- C-means (fuzzy) algorithm
- Agglomerative hierarchical clustering
All algorithms can be run in a fixed cluster mode and for a best estimate, i.e. when the number of clusters is not a priory given but is found after estimation of the cluster compactness. The data points can be defined in multidimensional space.
Data clustering is based on jMinHep package. You can run this in a completely stand-alone mode, without DataMelt. DataMelt integrates this Java program and enable Java scripting.
Using GUI
The easiest approach is to run a GUI editor to perform clustering. In the example below, we create several clusters in 3D and then passed the data holder to a GUI for clustering analysis:
from java.util import Random from jminhep.cluster import * from jhplot import * # create data for analysis data = DataHolder("Example") # fill 3D data with Gaussian random numbers rand = Random() for i in range(100): a =[] a.append( 10*rand.nextGaussian() ) a.append( 2*rand.nextGaussian()+1 ) a.append( 10*rand.nextGaussian()+3 ) data.add( DataPoint(a) ) # start jMinHEP GUI c1=HCluster(data)
This brings up a GUI editor which will run a selected algorithm:
Data clustering in JSAT
The package jsat.clustering/ provides another Java classes for data clustering. Here is an example that uses IRIS input data, runs multiple k-means algorithms, prints scores. Then it shows the clusters on a canvas:
from java.io import File from jsat import ARFFLoader,DataSet from jsat.classifiers import ClassificationDataSet from jsat.clustering import Clusterer from jsat.clustering.kmeans import KMeans,GMeans,HamerlyKMeans,KMeansPDN,XMeans,XMeans from jsat.clustering.evaluation import NormalizedMutualInformation from java.util.stream import IntStream print "Download iris_org.arff" from jhplot import * print Web.get("https://datamelt.org/examples/data/iris_org.arff") fi=File("iris_org.arff") dataSet = ARFFLoader.loadArffFile(fi) # We specify '0' as the class we would like to make the target class. data = ClassificationDataSet(dataSet, 0) """ We will use the NMI as our evaluation criteria. It compares the clustering results with the class labels. The class labels aren't necessarily the best ground truth for clusters. In fact, how to properly evaluate clustering algorithms is a very open question! But this is a commonly used method. he ClusterEvaluation interface dictates that values near 0 are better, and larger values are worse. NMI is usually the opposite, but obeys the interface. Read the NMI's Javadoc for more details. """ evaluator = NormalizedMutualInformation(); """ We will use a normal k-means algorithm to do clustering when we specify the number of clusters we want. JSAT implements a number of different algorithms that all solve the k-means problem, and are better in different scenarios. This one is likely to be the best for most users. """ simpleKMeans = HamerlyKMeans() from jarray import zeros clusteringResults = zeros(data.getSampleSize(), "i") # try different algorithms now.. def evaluate(methodsToEval): methodsToEval.cluster(data, clusteringResults) kFound = IntStream.of(clusteringResults).max().getAsInt()+1 print methodsToEval.toString(), " found=",kFound, " -> ",evaluator.evaluate(clusteringResults, data) evaluate(KMeansPDN()) evaluate(XMeans()) evaluate(GMeans()) print "Run k-means with a specific value of k, and keep track of cluster assignments" print "Row evaluate the cluster assignments and print a score.." print simpleKMeans.toString() for k in range(2,7): clusteringResults = simpleKMeans.cluster(data, k, clusteringResults); print "k=",k," score=",evaluator.evaluate(clusteringResults, data) print "Running for 3 clusters (optimal):" clusteringResults = simpleKMeans.cluster(data, 3, clusteringResults) clusters=clusteringResults.tolist() print "Cluster assignments=", clusters c1 = SPlot() c1.visible() c1.setAutoRange() c1.setMarksStyle('various') c1.setNameX('X') c1.setNameY('Y') for i in range(data.getSampleSize()): dataPoint = data.getDataPoint(i) category = data.getDataPointCategory(i) # get category vec=dataPoint.getNumericalValues(); c1.addPoint(clusters[i],vec.get(0),vec.get(1),1) if (i%10==0): c1.update() print i,dataPoint,category," Nr cluster=",clusters[i]
The canvas with the output (shown in 2D, for index 0 and 1) is:
Note that IRIS data are multidimensional data and cannot be easily visualized.
Using Jython scripts
Alternatively, one can run any clustering algorithm in batch mode without GUI. You can use Java, or any scripting programming language.
We show below a code which creates a data sample in 3D and then runs several clustering algorithms in one go. You can optionally print positions of the clusters and membership of the data points. The following modes will be used:
- K-means algorithm fixed cluster mode with single seed event
- K-means algorithm for multiple iterations
- K-means clustering using exchange method for best estimate
- K-means clustering using exchange method
- Hierarchical clustering algorithm
- Hierarchical clustering algorithm, best estimate
The following modes are available:
- 111 - standard k-means with single seed
- 112 - kmeans algorithm for multiple iterations
- 113 - k-means in exchange mode
- 114 - k-means multiple pass
- 121 - Hierarchical -- standard mode
- 122 - Hierarchical clustering algorithm, best estimate
- 131 - fuzzy
- 132 - fuzzy best estimate
from java.awt import Color from java.util import Random from jminhep.cluster import * # create a data holder data = DataHolder("Example") # fill 3D data with Gaussian random numbers rand = Random() for i in range(100): a =[] a.append( 10*rand.nextGaussian() ) a.append( 2*rand.nextGaussian()+1 ) a.append( 10*rand.nextGaussian()+3 ) data.add( DataPoint(a) ) del a # show the data # HTable(data) # Print data # data.print() # initialte partitioner pat = Partition(data); # set mode pat.set(3, 0.001, 1.7, 1000); # probability for membership (only for Fuzzy algorithm) pat.setProbab(0.68) # define types of cluster analysis mode =[] mode.append(111) mode.append(112) mode.append(113) mode.append(114) mode.append(121) mode.append(122) # mode.append(131) # mode.append(132) for i in range(len(mode)): print "test=",i pat.run(mode[i]) print "algorithm: " + pat.getName() print "Compactness: " + str(pat.getCompactness()) print "No of final clusters: " + str(pat.getNclusters()) Centers = pat.getCenters() # Centers.Print()
The output of the above script is shown below:
test 0 algorithm: kmeans algorithm fixed cluster mode with single seed event Compactness: 1.98271418832 No of final clusters: 3 test= 1 algorithm: kmeans algorithm for multiple iterations Compactness: 1.31227526642 No of final clusters: 3 test= 2 algorithm: K-means clustering using exchange method for best estimate Compactness: 1.35529140568 No of final clusters: 5 test= 3 algorithm: K-means clustering using exchange method Compactness: 1.35529140568 No of final clusters: 5 test= 4 algorithm: Hierarchical clustering algorithm Compactness: 1.41987639705 No of final clusters: 5 test= 5 algorithm: Hierarchical clustering algorithm, best estimate Compactness: 1.20134128248 No of final clusters: 6
You can print centers of clusters as:
Centers = pat.getCenters() for i in range(Centers.getSize()): g=Centers.getRow(i) n=g.getDimension() print i, " ", g.getAttribute(0),g.getAttribute(1),g.getAttribute(2)
Where "Centers" are "DataHolder" container in the above example.