// Copyright (C) 2011 Zeno Gantner, Chris Newell
//
// This file is part of MyMediaLite.
//
// MyMediaLite is free software: you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation, either version 3 of the License, or
// (at your option) any later version.
//
// MyMediaLite is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with MyMediaLite. If not, see <http://www.gnu.org/licenses/>.
package org.mymedialite.ratingprediction;
import it.unimi.dsi.fastutil.doubles.DoubleArrayList;
import it.unimi.dsi.fastutil.doubles.DoubleList;
import it.unimi.dsi.fastutil.ints.IntArrayList;
import it.unimi.dsi.fastutil.ints.IntList;
import java.io.BufferedReader;
import java.io.IOException;
import java.io.PrintWriter;
import java.util.ArrayList;
import java.util.List;
import org.mymedialite.IIterativeModel;
import org.mymedialite.datatype.IMatrix;
import org.mymedialite.datatype.Matrix;
import org.mymedialite.eval.Ratings;
import org.mymedialite.io.IMatrixExtensions;
import org.mymedialite.io.Model;
import org.mymedialite.io.VectorExtensions;
import org.mymedialite.util.Random;
/**
* Co-clustering for rating prediction.
*
* Literature:
*
* Thomas George, Srujana Merugu
* A Scalable Collaborative Filtering Framework based on Co-clustering.
* ICDM 2005.
* http://hercules.ece.utexas.edu/~srujana/papers/icdm05.pdf
*
* This recommender does NOT support incremental updates.
* @version 2.03
*/
public class CoClustering extends RatingPredictor implements IIterativeModel {
private static final String VERSION = "2.03";
/**
* Random number generator.
*/
org.mymedialite.util.Random random;
IntList user_clustering;
IntList item_clustering;
DoubleList user_averages;
DoubleList item_averages;
IntList user_counts;
IntList item_counts;
DoubleList user_cluster_averages;
DoubleList item_cluster_averages;
IMatrix<Double> cocluster_averages;
double global_average;
/**
* The number of user clusters.
*/
public int numUserClusters;
/**
* The number of item clusters.
*/
public int numItemClusters;
/**
* The maximum number of iterations.
* If the algorithm converges to a stable solution, it will terminate earlier.
*/
public int numIter;
/**
* Default constructor.
*/
public CoClustering() {
numUserClusters = 3;
numItemClusters = 3;
numIter = 30;
}
@Override
public void setNumIter(int numIter) {
this.numIter = numIter;
}
@Override
public int getNumIter() {
return numIter;
}
public void initModel() {
this.user_clustering = new IntArrayList();
user_clustering.size(maxUserID + 1);
this.item_clustering = new IntArrayList();
item_clustering.size(maxItemID + 1);
this.user_cluster_averages = new DoubleArrayList();
user_cluster_averages.size(numUserClusters);
this.item_cluster_averages = new DoubleArrayList();
item_cluster_averages.size(numItemClusters);
this.cocluster_averages = new Matrix<Double>(numUserClusters, numItemClusters, 0.0);
}
boolean iterateCheckModified() {
boolean clustering_modified = false;
computeClusterAverages();
// Dimension users
for (int u = 0; u < maxUserID; u++)
if (findOptimalUserClustering(u))
clustering_modified = true;
// Dimension items
for (int i = 0; i < maxItemID; i++)
if (findOptimalItemClustering(i))
clustering_modified = true;
return clustering_modified;
}
/**
*
*/
public void iterate() {
iterateCheckModified();
}
/**
*
*/
public void train() {
random = Random.getInstance();
initModel();
for (int i = 0; i < user_clustering.size(); i++)
user_clustering.set(i, random.nextInt(numUserClusters));
for (int i = 0; i < item_clustering.size(); i++)
item_clustering.set(i, random.nextInt(numItemClusters));
computeAverages();
for (int i = 0; i < numIter; i++)
if (!iterateCheckModified())
break;
}
/**
*
*/
public double predict(int u, int i) {
if (u > maxUserID && i > maxItemID)
return global_average;
if (u > maxUserID)
return item_cluster_averages.get(item_clustering.get(i));
if (i > maxItemID)
return user_cluster_averages.get(user_clustering.get(u));
double prediction = predict(u, i, user_clustering.get(u), item_clustering.get(i));
if (prediction < minRating)
return minRating;
if (prediction > maxRating)
return maxRating;
return prediction;
}
double predict(int u, int i, int uc, int ic) {
return cocluster_averages.get(uc, ic)
+ user_averages.get(u)
- user_cluster_averages.get(uc)
+ item_averages.get(i)
- item_cluster_averages.get(ic);
}
boolean findOptimalUserClustering(int user_id) {
boolean modified = false;
double[] errors = new double[numUserClusters];
for (int uc = 0; uc < numUserClusters; uc++)
for (int index : ratings.byUser().get(user_id)) {
int item_id = ratings.items().get(index);
double rating = ratings.get(index);
errors[uc] += Math.pow(rating - predict(user_id, item_id, uc, item_clustering.get(item_id)), 2);
}
int minimum_index = getMinimumIndex(errors, user_clustering.get(user_id));
if (minimum_index != user_clustering.get(user_id)) {
user_clustering.set(user_id, minimum_index);
modified = true;
}
return modified;
}
boolean findOptimalItemClustering(int item_id) {
boolean modified = false;
double[] errors = new double[numItemClusters];
for (int ic = 0; ic < numItemClusters; ic++)
for (int index : ratings.byItem().get(item_id)) {
int user_id = ratings.users().get(index);
double rating = ratings.get(index);
errors[ic] += Math.pow(rating - predict(user_id, item_id, user_clustering.get(user_id), ic), 2);
}
int minimum_index = getMinimumIndex(errors, item_clustering.get(item_id));
if (minimum_index != item_clustering.get(item_id)) {
item_clustering.set(item_id, minimum_index);
modified = true;
}
return modified;
}
void computeAverages() {
double[] user_sums = new double[maxUserID + 1];
double[] item_sums = new double[maxItemID + 1];
double sum = 0;
this.user_counts = new IntArrayList();
user_counts.size(maxUserID + 1);
this.item_counts = new IntArrayList();
item_counts.size(maxItemID + 1);
for (int i = 0; i < ratings.size(); i++) {
int user_id = ratings.users().get(i);
int item_id = ratings.items().get(i);
double rating = ratings.get(i);
user_sums[user_id] += rating;
item_sums[item_id] += rating;
sum += rating;
user_counts.set(user_id, user_counts.get(user_id) + 1);
item_counts.set(item_id, item_counts.get(item_id) + 1);
}
this.global_average = sum / ratings.size();
this.user_averages = new DoubleArrayList();
user_averages.size(maxUserID + 1);
for (int u = 0; u <= maxUserID; u++)
if (user_counts.get(u) > 0)
user_averages.set(u, user_sums[u] / user_counts.get(u));
else
user_averages.set(u, global_average);
this.item_averages = new DoubleArrayList();
item_averages.size(maxItemID + 1);
for (int i = 0; i <= maxItemID; i++)
if (item_counts.get(i) > 0)
item_averages.set(i, item_sums[i] / item_counts.get(i));
else
item_averages.set(i, global_average);
}
void computeClusterAverages() {
int[] user_cluster_counts = new int[numUserClusters];
int[] item_cluster_counts = new int[numItemClusters];
int[][] cocluster_counts = new int[numUserClusters][numItemClusters];
for (int i = 0; i < ratings.size(); i++) {
int user_id = ratings.users().get(i);
int item_id = ratings.items().get(i);
double rating = ratings.get(i);
user_cluster_averages.set(user_clustering.getInt(user_id), user_cluster_averages.getDouble(user_clustering.get(user_id)) + rating);
item_cluster_averages.set(item_clustering.getInt(item_id), item_cluster_averages.getDouble(item_clustering.get(item_id)) + rating);
cocluster_averages.set(user_clustering.get(user_id), item_clustering.get(item_id), cocluster_averages.get(user_clustering.get(user_id), item_clustering.get(item_id)) + rating);
user_cluster_counts[user_clustering.get(user_id)]++;
item_cluster_counts[item_clustering.get(item_id)]++;
cocluster_counts[user_clustering.get(user_id)][ item_clustering.get(item_id)]++;
}
for (int i = 0; i < numUserClusters; i++)
if (user_cluster_counts[i] > 0)
user_cluster_averages.set(i, user_cluster_averages.get(i) / user_cluster_counts[i]);
else
user_cluster_averages.set(i, global_average);
for (int i = 0; i < numItemClusters; i++)
if (item_cluster_counts[i] > 0)
item_cluster_averages.set(i, item_cluster_averages.get(i) / item_cluster_counts[i]);
else
item_cluster_averages.set(i, global_average);
for (int i = 0; i < numUserClusters; i++)
for (int j = 0; j < numItemClusters; j++)
if (cocluster_counts[i][j] > 0)
cocluster_averages.set(i, j, cocluster_averages.get(i, j) / cocluster_counts[i][j]);
else
cocluster_averages.set(i, j, global_average);
}
int getMinimumIndex(double[] array, int default_index) {
int minimumIndex = default_index;
for (int i = 0; i < array.length; i++)
if (array[i] < array[minimumIndex])
minimumIndex = i;
return minimumIndex;
}
/**
*/
public void saveModel(String filename) throws IOException {
PrintWriter writer = Model.getWriter(filename, this.getClass(), VERSION);
saveModel(writer);
writer.flush();
writer.close();
}
@Override
public void saveModel(PrintWriter writer) throws IOException {
VectorExtensions.writeVector(writer, user_clustering);
VectorExtensions.writeVector(writer, item_clustering);
writer.println(Double.toString(global_average));
VectorExtensions.writeVector(writer, user_averages);
VectorExtensions.writeVector(writer, item_averages);
VectorExtensions.writeVector(writer, user_cluster_averages);
VectorExtensions.writeVector(writer, item_cluster_averages);
IMatrixExtensions.writeMatrix(writer, cocluster_averages);
}
@Override
public void loadModel(String filename) throws IOException {
BufferedReader reader = Model.getReader(filename, this.getClass());
loadModel(reader);
reader.close();
}
@Override
public void loadModel(BufferedReader reader) throws IOException {
IntList user_clustering = VectorExtensions.readIntVector(reader);
IntList item_clustering = VectorExtensions.readIntVector(reader);
double global_average = Double.parseDouble(reader.readLine());
DoubleList user_averages = VectorExtensions.readVector(reader);
DoubleList item_averages = VectorExtensions.readVector(reader);
DoubleList user_cluster_averages = VectorExtensions.readVector(reader);
DoubleList item_cluster_averages = VectorExtensions.readVector(reader);
IMatrix<Double> cocluster_averages = IMatrixExtensions.readDoubleMatrix(reader, new Matrix<Double>(0, 0));
reader.close();
int num_user_clusters = user_cluster_averages.size();
int num_item_clusters = item_cluster_averages.size();
// Adjust maximum IDs
this.maxUserID = user_clustering.size() - 1;
this.maxItemID = item_clustering.size() - 1;
// Adjust hyperparameters
if (this.numUserClusters != num_user_clusters) {
System.err.println("Set num_user_clusters to " + num_user_clusters);
this.numUserClusters = num_user_clusters;
}
if (this.numItemClusters != num_item_clusters) {
System.err.println("Set num_item_clusters to " + num_item_clusters);
this.numItemClusters = num_item_clusters;
}
// Assign model
this.global_average = global_average;
this.user_cluster_averages = user_cluster_averages;
this.item_cluster_averages = item_cluster_averages;
this.cocluster_averages = cocluster_averages;
this.user_averages = user_averages;
this.item_averages = item_averages;
this.user_clustering = user_clustering;
this.item_clustering = item_clustering;
}
/**
*
*/
public double computeLoss() {
return Ratings.evaluate(this, ratings).get("RMSE");
}
/**
*
*/
public String toString() {
return
this.getClass().getName()
+ " numUserClusters=" + numUserClusters
+ " numItemClusters=" + numItemClusters
+ " numIter=" + numIter;
}
}