protected static void golf(InstanceHolder trainingSet, InstanceHolder testSet) throws Exception {
int numatts = 3;
int fromAtt = 8;
InstanceHolder filteredTrain = new InstanceHolder(trainingSet.getNumberOfClasses(), numatts);
InstanceHolder filteredTest = new InstanceHolder(testSet.getNumberOfClasses(), numatts);
for (int i = 0; i < trainingSet.size(); i++) {
SparseVector instance = new SparseVector(numatts);
for (int j = 0; j < numatts; j++) {
instance.put(j, trainingSet.getInstance(i).get(j + fromAtt));
}
filteredTrain.add(instance, trainingSet.getLabel(i));
}
for (int i = 0; i < testSet.size(); i++) {
SparseVector instance = new SparseVector(numatts);
for (int j = 0; j < numatts; j++) {
instance.put(j, testSet.getInstance(i).get(j + fromAtt));
}
filteredTest.add(instance, testSet.getLabel(i));
}
KMeans kMeans = new KMeans(numOfClusters, 250.0);
Random r = new Random(seed);
for (int i = 0; i < 2 * filteredTrain.size(); i++) {
int index = r.nextInt(filteredTrain.size());
kMeans.update(filteredTrain.getInstance(index), filteredTrain.getLabel(index));
}
Model[] models = new Model[numOfClusters];
for (int i = 0; i < numOfClusters; i++) {
models[i] = (Model) Class.forName(modelName).newInstance();
models[i].init(prefix);
models[i].setNumberOfClasses((int) numberOfRatings);
}
for (int i = 0; i < 10 * trainingSet.size(); i++) {
int index = r.nextInt(trainingSet.size());
int clusterID = (int) kMeans.predict(filteredTrain.getInstance(index));
models[clusterID].update(trainingSet.getInstance(index), trainingSet.getLabel(index));
// if (i%100 == 0) System.out.println(i + "\t" + evaluate(models[clusterID], trainingSet) +
// "\t" + evaluate(models[clusterID], testSet));// + "\t" + models[clusterID]);
}
double MAError = 0.0;
double RMSError = 0.0;
for (int i = 0; i < testSet.size(); i++) {
int clusterID = (int) kMeans.predict(filteredTest.getInstance(i));
double predicted = models[clusterID].predict(testSet.getInstance(i));
double expected = testSet.getLabel(i);
// System.out.println(expected + "\t" + predicted);
MAError += Math.abs(expected - predicted);
RMSError += Math.pow(((expected - predicted) / divErr), 2);
}
MAError /= testSet.size() * divErr;
RMSError /= testSet.size();
RMSError = Math.sqrt(RMSError);
System.out.println("GoLF MAE: " + MAError);
System.out.println("GoLF RMSE: " + RMSError);
/*for (int i = 0; i < numOfClusters; i++) {
System.out.println(models[i]);
}*/
}
protected static void computeUserSpecificFeatureVectorPart(
SparseVector userRatings, double[] featureVectorA) {
// compute the averaged rating of the user
double userAvgRating = 0.0, userNumberOfRatings = 0.0;
for (VectorEntry itemRating : userRatings) {
userAvgRating += itemRating.value;
userNumberOfRatings++;
}
userAvgRating =
(userNumberOfRatings > 0.0)
? userAvgRating / userNumberOfRatings
: (numberOfRatings + 1.0)
/ 2.0; // when no rating for a user, use the mean of rating range
featureVectorA[0] = userAvgRating / numberOfRatings;
featureVectorA[1] = userNumberOfRatings / numberOfAllUsers;
double likeSize = 0.0, disLikeSize = 0.0;
// extract features
for (VectorEntry itemRating : userRatings) {
int itemID = itemRating.index;
double rating = itemRating.value; // classLabel
double itemPop = getNumberOfUsers(itemID) / numberOfAllUsers;
double itemAvg = getAverageRating(itemID) / numberOfRatings;
// general
featureVectorA[2] += itemPop;
featureVectorA[3] += itemAvg;
// like
if (rating > userAvgRating) {
featureVectorA[4] += itemPop;
featureVectorA[5] += itemAvg;
likeSize++;
}
// dislike
if (rating < userAvgRating) {
featureVectorA[6] += itemPop;
featureVectorA[7] += itemAvg;
disLikeSize++;
}
}
// normalize features
featureVectorA[2] =
(userRatings.size() > 0) ? featureVectorA[2] / (double) userRatings.size() : 0.0;
featureVectorA[3] =
(userRatings.size() > 0) ? featureVectorA[3] / (double) userRatings.size() : 0.0;
featureVectorA[4] = (likeSize > 0.0) ? featureVectorA[4] / likeSize : 0.0;
featureVectorA[5] = (likeSize > 0.0) ? featureVectorA[5] / likeSize : 0.0;
featureVectorA[6] = (disLikeSize > 0.0) ? featureVectorA[6] / disLikeSize : 0.0;
featureVectorA[7] = (disLikeSize > 0.0) ? featureVectorA[7] / disLikeSize : 0.0;
}
/**
* This method parses the given file into collections of instances and corresponding class labels.
*
* @param file the file that has to be parsed
* @throws IOException if file reading error occurs.
*/
protected InstanceHolder parseFile(final File file) throws IOException {
// throw exception if the file does not exist or null
if (file == null || !file.exists()) {
throw new RuntimeException("The file \"" + file.toString() + "\" is null or does not exist!");
}
// InstanceHolder holder = new InstanceHolder();
Vector<SparseVector> instances = new Vector<SparseVector>();
Vector<Double> labels = new Vector<Double>();
BufferedReader br = new BufferedReader(new FileReader(file));
int numberOfClasses = -1;
int numberOfFeatures = -1;
String line;
String[] split;
int c = 0;
double rate;
int userId;
int itemId;
SparseVector instance;
while ((line = br.readLine()) != null) {
c++;
// eliminating empty and comment lines
if (line.length() == 0 || line.startsWith("#")) {
continue;
}
// eliminating comments and white spaces from the endings of the line
line = line.replaceAll("#.*", "").trim();
// splitting line at white spaces and at colons
split = line.split("\\s");
// throwing exception if the line is invalid (= has even number of tokens, since
// a valid line has a class label and pairs of indices and corresponding values)
if (split.length != 3) {
throw new RuntimeException(
"The file \"" + file.toString() + "\" has invalid structure at line " + c);
}
userId = Integer.parseInt(split[0]) - 1;
itemId = Integer.parseInt(split[1]) - 1;
rate = Double.parseDouble(split[2]);
/*if (numberOfClasses != Integer.MAX_VALUE && (rate <= 0.0 || rate != (int)rate)) {
// not a regression problem => the label has to be an integer which is greater or equal than 0
throw new RuntimeException("The rate value has to be integer and greater than 0, line " + c);
}*/
if (userId < 0.0) {
throw new RuntimeException(
"The user ID has to be integer and greater than or equal to 0, line " + c);
}
if (itemId < 0.0) {
throw new RuntimeException(
"The item ID has to be integer and greater than or equal to 0, line " + c);
}
if (rate > numberOfClasses) {
numberOfClasses = (int) rate;
}
if (itemId >= numberOfFeatures) {
numberOfFeatures = itemId + 1;
}
if (instances.size() <= userId) {
for (int i = instances.size(); i <= userId; i++) {
instances.add(new SparseVector(1));
labels.add(0.0);
}
}
instance = instances.get(userId);
instance.put(itemId, rate);
labels.set(userId, (double) userId);
}
br.close();
return new InstanceHolder(
instances,
labels,
(numberOfClasses == 1) ? 0 : numberOfClasses,
numberOfFeatures); // 1-> indicating clustering
}
