/**
* Evaluates a clusterer with the options given in an array of strings. It takes the string
* indicated by "-t" as training file, the string indicated by "-T" as test file. If the test file
* is missing, a stratified ten-fold cross-validation is performed (distribution clusterers only).
* Using "-x" you can change the number of folds to be used, and using "-s" the random seed. If
* the "-p" option is present it outputs the classification for each test instance. If you provide
* the name of an object file using "-l", a clusterer will be loaded from the given file. If you
* provide the name of an object file using "-d", the clusterer built from the training data will
* be saved to the given file.
*
* @param clusterer machine learning clusterer
* @param options the array of string containing the options
* @exception Exception if model could not be evaluated successfully
* @return a string describing the results
*/
public static String evaluateClusterer(Clusterer clusterer, String[] options) throws Exception {
int seed = 1, folds = 10;
boolean doXval = false;
Instances train = null;
Instances test = null;
Random random;
String trainFileName,
testFileName,
seedString,
foldsString,
objectInputFileName,
objectOutputFileName,
attributeRangeString;
String[] savedOptions = null;
boolean printClusterAssignments = false;
Range attributesToOutput = null;
ObjectInputStream objectInputStream = null;
ObjectOutputStream objectOutputStream = null;
StringBuffer text = new StringBuffer();
int theClass = -1; // class based evaluation of clustering
try {
if (Utils.getFlag('h', options)) {
throw new Exception("Help requested.");
}
// Get basic options (options the same for all clusterers
// printClusterAssignments = Utils.getFlag('p', options);
objectInputFileName = Utils.getOption('l', options);
objectOutputFileName = Utils.getOption('d', options);
trainFileName = Utils.getOption('t', options);
testFileName = Utils.getOption('T', options);
// Check -p option
try {
attributeRangeString = Utils.getOption('p', options);
} catch (Exception e) {
throw new Exception(
e.getMessage()
+ "\nNOTE: the -p option has changed. "
+ "It now expects a parameter specifying a range of attributes "
+ "to list with the predictions. Use '-p 0' for none.");
}
if (attributeRangeString.length() != 0) {
printClusterAssignments = true;
if (!attributeRangeString.equals("0")) attributesToOutput = new Range(attributeRangeString);
}
if (trainFileName.length() == 0) {
if (objectInputFileName.length() == 0) {
throw new Exception("No training file and no object " + "input file given.");
}
if (testFileName.length() == 0) {
throw new Exception("No training file and no test file given.");
}
} else {
if ((objectInputFileName.length() != 0) && (printClusterAssignments == false)) {
throw new Exception("Can't use both train and model file " + "unless -p specified.");
}
}
seedString = Utils.getOption('s', options);
if (seedString.length() != 0) {
seed = Integer.parseInt(seedString);
}
foldsString = Utils.getOption('x', options);
if (foldsString.length() != 0) {
folds = Integer.parseInt(foldsString);
doXval = true;
}
} catch (Exception e) {
throw new Exception('\n' + e.getMessage() + makeOptionString(clusterer));
}
try {
if (trainFileName.length() != 0) {
train = new Instances(new BufferedReader(new FileReader(trainFileName)));
String classString = Utils.getOption('c', options);
if (classString.length() != 0) {
if (classString.compareTo("last") == 0) {
theClass = train.numAttributes();
} else if (classString.compareTo("first") == 0) {
theClass = 1;
} else {
theClass = Integer.parseInt(classString);
}
if (doXval || testFileName.length() != 0) {
throw new Exception("Can only do class based evaluation on the " + "training data");
}
if (objectInputFileName.length() != 0) {
throw new Exception("Can't load a clusterer and do class based " + "evaluation");
}
}
if (theClass != -1) {
if (theClass < 1 || theClass > train.numAttributes()) {
throw new Exception("Class is out of range!");
}
if (!train.attribute(theClass - 1).isNominal()) {
throw new Exception("Class must be nominal!");
}
train.setClassIndex(theClass - 1);
}
}
if (objectInputFileName.length() != 0) {
objectInputStream = new ObjectInputStream(new FileInputStream(objectInputFileName));
}
if (objectOutputFileName.length() != 0) {
objectOutputStream = new ObjectOutputStream(new FileOutputStream(objectOutputFileName));
}
} catch (Exception e) {
throw new Exception("ClusterEvaluation: " + e.getMessage() + '.');
}
// Save options
if (options != null) {
savedOptions = new String[options.length];
System.arraycopy(options, 0, savedOptions, 0, options.length);
}
if (objectInputFileName.length() != 0) {
Utils.checkForRemainingOptions(options);
}
// Set options for clusterer
if (clusterer instanceof OptionHandler) {
((OptionHandler) clusterer).setOptions(options);
}
Utils.checkForRemainingOptions(options);
if (objectInputFileName.length() != 0) {
// Load the clusterer from file
clusterer = (Clusterer) objectInputStream.readObject();
objectInputStream.close();
} else {
// Build the clusterer if no object file provided
if (theClass == -1) {
clusterer.buildClusterer(train);
} else {
Remove removeClass = new Remove();
removeClass.setAttributeIndices("" + theClass);
removeClass.setInvertSelection(false);
removeClass.setInputFormat(train);
Instances clusterTrain = Filter.useFilter(train, removeClass);
clusterer.buildClusterer(clusterTrain);
ClusterEvaluation ce = new ClusterEvaluation();
ce.setClusterer(clusterer);
ce.evaluateClusterer(train);
return "\n\n=== Clustering stats for training data ===\n\n" + ce.clusterResultsToString();
}
}
/* Output cluster predictions only (for the test data if specified,
otherwise for the training data */
if (printClusterAssignments) {
return printClusterings(clusterer, train, testFileName, attributesToOutput);
}
text.append(clusterer.toString());
text.append(
"\n\n=== Clustering stats for training data ===\n\n"
+ printClusterStats(clusterer, trainFileName));
if (testFileName.length() != 0) {
text.append(
"\n\n=== Clustering stats for testing data ===\n\n"
+ printClusterStats(clusterer, testFileName));
}
if ((clusterer instanceof DensityBasedClusterer)
&& (doXval == true)
&& (testFileName.length() == 0)
&& (objectInputFileName.length() == 0)) {
// cross validate the log likelihood on the training data
random = new Random(seed);
random.setSeed(seed);
train.randomize(random);
text.append(
crossValidateModel(clusterer.getClass().getName(), train, folds, savedOptions, random));
}
// Save the clusterer if an object output file is provided
if (objectOutputFileName.length() != 0) {
objectOutputStream.writeObject(clusterer);
objectOutputStream.flush();
objectOutputStream.close();
}
return text.toString();
}
/**
* Evaluate the clusterer on a set of instances. Calculates clustering statistics and stores
* cluster assigments for the instances in m_clusterAssignments
*
* @param test the set of instances to cluster
* @exception Exception if something goes wrong
*/
public void evaluateClusterer(Instances test) throws Exception {
int i = 0;
int cnum;
double loglk = 0.0;
double[] dist;
double temp;
int cc = m_Clusterer.numberOfClusters();
m_numClusters = cc;
int numInstFieldWidth = (int) ((Math.log(test.numInstances()) / Math.log(10)) + 1);
double[] instanceStats = new double[cc];
m_clusterAssignments = new double[test.numInstances()];
Instances testCopy = test;
boolean hasClass = (testCopy.classIndex() >= 0);
int unclusteredInstances = 0;
// If class is set then do class based evaluation as well
if (hasClass) {
if (testCopy.classAttribute().isNumeric()) {
throw new Exception("ClusterEvaluation: Class must be nominal!");
}
Remove removeClass = new Remove();
removeClass.setAttributeIndices("" + (testCopy.classIndex() + 1));
removeClass.setInvertSelection(false);
removeClass.setInputFormat(testCopy);
testCopy = Filter.useFilter(testCopy, removeClass);
}
for (i = 0; i < testCopy.numInstances(); i++) {
cnum = -1;
try {
if (m_Clusterer instanceof DensityBasedClusterer) {
loglk +=
((DensityBasedClusterer) m_Clusterer).logDensityForInstance(testCopy.instance(i));
// temp = Utils.sum(dist);
// Utils.normalize(dist);
cnum = m_Clusterer.clusterInstance(testCopy.instance(i));
// Utils.maxIndex(dist);
m_clusterAssignments[i] = (double) cnum;
} else {
cnum = m_Clusterer.clusterInstance(testCopy.instance(i));
m_clusterAssignments[i] = (double) cnum;
}
} catch (Exception e) {
unclusteredInstances++;
}
if (cnum != -1) {
instanceStats[cnum]++;
}
}
/* // count the actual number of used clusters
int count = 0;
for (i = 0; i < cc; i++) {
if (instanceStats[i] > 0) {
count++;
}
}
if (count > 0) {
double [] tempStats = new double [count];
double [] map = new double [m_clusterAssignments.length];
count=0;
for (i=0;i<cc;i++) {
if (instanceStats[i] > 0) {
tempStats[count] = instanceStats[i];
map[i] = count;
count++;
}
}
instanceStats = tempStats;
cc = instanceStats.length;
for (i=0;i<m_clusterAssignments.length;i++) {
m_clusterAssignments[i] = map[(int)m_clusterAssignments[i]];
}
} */
double sum = Utils.sum(instanceStats);
loglk /= sum;
m_logL = loglk;
m_clusteringResults.append(m_Clusterer.toString());
m_clusteringResults.append("Clustered Instances\n\n");
int clustFieldWidth = (int) ((Math.log(cc) / Math.log(10)) + 1);
for (i = 0; i < cc; i++) {
if (instanceStats[i] > 0) {
m_clusteringResults.append(
Utils.doubleToString((double) i, clustFieldWidth, 0)
+ " "
+ Utils.doubleToString(instanceStats[i], numInstFieldWidth, 0)
+ " ("
+ Utils.doubleToString((instanceStats[i] / sum * 100.0), 3, 0)
+ "%)\n");
}
}
if (unclusteredInstances > 0) {
m_clusteringResults.append("\nUnclustered instances : " + unclusteredInstances);
}
if (m_Clusterer instanceof DensityBasedClusterer) {
m_clusteringResults.append("\n\nLog likelihood: " + Utils.doubleToString(loglk, 1, 5) + "\n");
}
if (hasClass) {
evaluateClustersWithRespectToClass(test);
}
}