/**
* Licensed to the Apache Software Foundation (ASF) under one or more
* contributor license agreements. See the NOTICE file distributed with
* this work for additional information regarding copyright ownership.
* The ASF licenses this file to You under the Apache License, Version 2.0
* (the "License"); you may not use this file except in compliance with
* the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.mahout.cf.taste.impl.recommender;
import java.util.Collection;
import java.util.Collections;
import java.util.Iterator;
import java.util.List;
import java.util.ListIterator;
import java.util.PriorityQueue;
import java.util.Queue;
import java.util.concurrent.Callable;
import com.google.common.collect.Lists;
import org.apache.mahout.cf.taste.common.Refreshable;
import org.apache.mahout.cf.taste.common.TasteException;
import org.apache.mahout.cf.taste.impl.common.FastByIDMap;
import org.apache.mahout.cf.taste.impl.common.FastIDSet;
import org.apache.mahout.cf.taste.impl.common.FullRunningAverage;
import org.apache.mahout.cf.taste.impl.common.LongPrimitiveIterator;
import org.apache.mahout.cf.taste.impl.common.RefreshHelper;
import org.apache.mahout.cf.taste.impl.common.RunningAverage;
import org.apache.mahout.cf.taste.model.DataModel;
import org.apache.mahout.cf.taste.recommender.ClusteringRecommender;
import org.apache.mahout.cf.taste.recommender.IDRescorer;
import org.apache.mahout.cf.taste.recommender.RecommendedItem;
import org.apache.mahout.common.RandomUtils;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import com.google.common.base.Preconditions;
/**
* <p>
* A {@link org.apache.mahout.cf.taste.recommender.Recommender} that clusters users, then determines the
* clusters' top recommendations. This implementation builds clusters by repeatedly merging clusters until
* only a certain number remain, meaning that each cluster is sort of a tree of other clusters.
* </p>
*
* <p>
* This {@link org.apache.mahout.cf.taste.recommender.Recommender} therefore has a few properties to note:
* </p>
* <ul>
* <li>For all users in a cluster, recommendations will be the same</li>
* <li>{@link #estimatePreference(long, long)} may well return {@link Double#NaN}; it does so when asked to
* estimate preference for an item for which no preference is expressed in the users in the cluster.</li>
* </ul>
*
* <p>
* This is an <em>experimental</em> implementation which tries to gain a lot of speed at the cost of accuracy
* in building clusters, compared to {@link TreeClusteringRecommender}. It will sometimes cluster two other
* clusters together that may not be the exact closest two clusters in existence. This may not affect the
* recommendation quality much, but it potentially speeds up the clustering process dramatically.
* </p>
*/
public final class TreeClusteringRecommender2 extends AbstractRecommender implements ClusteringRecommender {
private static final Logger log = LoggerFactory.getLogger(TreeClusteringRecommender2.class);
private static final int NUM_CLUSTER_RECS = 100;
private final ClusterSimilarity clusterSimilarity;
private final int numClusters;
private final double clusteringThreshold;
private final boolean clusteringByThreshold;
private FastByIDMap<List<RecommendedItem>> topRecsByUserID;
private FastIDSet[] allClusters;
private FastByIDMap<FastIDSet> clustersByUserID;
private final RefreshHelper refreshHelper;
/**
* @param dataModel
* {@link DataModel} which provides users
* @param clusterSimilarity
* {@link ClusterSimilarity} used to compute cluster similarity
* @param numClusters
* desired number of clusters to create
* @throws IllegalArgumentException
* if arguments are {@code null}, or {@code numClusters} is less than 2
*/
public TreeClusteringRecommender2(DataModel dataModel, ClusterSimilarity clusterSimilarity, int numClusters)
throws TasteException {
super(dataModel);
Preconditions.checkArgument(numClusters >= 2, "numClusters must be at least 2");
this.clusterSimilarity = Preconditions.checkNotNull(clusterSimilarity);
this.numClusters = numClusters;
this.clusteringThreshold = Double.NaN;
this.clusteringByThreshold = false;
this.refreshHelper = new RefreshHelper(new Callable<Object>() {
@Override
public Object call() throws TasteException {
buildClusters();
return null;
}
});
refreshHelper.addDependency(dataModel);
refreshHelper.addDependency(clusterSimilarity);
buildClusters();
}
/**
* @param dataModel
* {@link DataModel} which provides users
* @param clusterSimilarity
* {@link ClusterSimilarity} used to compute cluster
* similarity
* @param clusteringThreshold
* clustering similarity threshold; clusters will be aggregated into larger clusters until the next
* two nearest clusters' similarity drops below this threshold
* @throws IllegalArgumentException
* if arguments are {@code null}, or {@code clusteringThreshold} is {@link Double#NaN}
*/
public TreeClusteringRecommender2(DataModel dataModel,
ClusterSimilarity clusterSimilarity,
double clusteringThreshold) throws TasteException {
super(dataModel);
Preconditions.checkArgument(!Double.isNaN(clusteringThreshold), "clusteringThreshold must not be NaN");
this.clusterSimilarity = Preconditions.checkNotNull(clusterSimilarity);
this.numClusters = Integer.MIN_VALUE;
this.clusteringThreshold = clusteringThreshold;
this.clusteringByThreshold = true;
this.refreshHelper = new RefreshHelper(new Callable<Object>() {
@Override
public Object call() throws TasteException {
buildClusters();
return null;
}
});
refreshHelper.addDependency(dataModel);
refreshHelper.addDependency(clusterSimilarity);
buildClusters();
}
@Override
public List<RecommendedItem> recommend(long userID, int howMany, IDRescorer rescorer) throws TasteException {
Preconditions.checkArgument(howMany >= 1, "howMany must be at least 1");
buildClusters();
log.debug("Recommending items for user ID '{}'", userID);
List<RecommendedItem> recommended = topRecsByUserID.get(userID);
if (recommended == null) {
return Collections.emptyList();
}
DataModel dataModel = getDataModel();
List<RecommendedItem> rescored = Lists.newArrayListWithCapacity(recommended.size());
// Only add items the user doesn't already have a preference for.
// And that the rescorer doesn't "reject".
for (RecommendedItem recommendedItem : recommended) {
long itemID = recommendedItem.getItemID();
if (rescorer != null && rescorer.isFiltered(itemID)) {
continue;
}
if (dataModel.getPreferenceValue(userID, itemID) == null
&& (rescorer == null || !Double.isNaN(rescorer.rescore(itemID, recommendedItem.getValue())))) {
rescored.add(recommendedItem);
}
}
Collections.sort(rescored, new ByRescoreComparator(rescorer));
return rescored;
}
@Override
public float estimatePreference(long userID, long itemID) throws TasteException {
Float actualPref = getDataModel().getPreferenceValue(userID, itemID);
if (actualPref != null) {
return actualPref;
}
buildClusters();
List<RecommendedItem> topRecsForUser = topRecsByUserID.get(userID);
if (topRecsForUser != null) {
for (RecommendedItem item : topRecsForUser) {
if (itemID == item.getItemID()) {
return item.getValue();
}
}
}
// Hmm, we have no idea. The item is not in the user's cluster
return Float.NaN;
}
@Override
public FastIDSet getCluster(long userID) throws TasteException {
buildClusters();
FastIDSet cluster = clustersByUserID.get(userID);
return cluster == null ? new FastIDSet() : cluster;
}
@Override
public FastIDSet[] getClusters() throws TasteException {
buildClusters();
return allClusters;
}
private static final class ClusterClusterPair implements Comparable<ClusterClusterPair> {
private final FastIDSet cluster1;
private final FastIDSet cluster2;
private final double similarity;
private ClusterClusterPair(FastIDSet cluster1, FastIDSet cluster2, double similarity) {
this.cluster1 = cluster1;
this.cluster2 = cluster2;
this.similarity = similarity;
}
FastIDSet getCluster1() {
return cluster1;
}
FastIDSet getCluster2() {
return cluster2;
}
double getSimilarity() {
return similarity;
}
@Override
public int hashCode() {
return cluster1.hashCode() ^ cluster2.hashCode() ^ RandomUtils.hashDouble(similarity);
}
@Override
public boolean equals(Object o) {
if (!(o instanceof ClusterClusterPair)) {
return false;
}
ClusterClusterPair other = (ClusterClusterPair) o;
return cluster1.equals(other.getCluster1())
&& cluster2.equals(other.getCluster2())
&& similarity == other.getSimilarity();
}
@Override
public int compareTo(ClusterClusterPair other) {
double otherSimilarity = other.getSimilarity();
if (similarity > otherSimilarity) {
return -1;
} else if (similarity < otherSimilarity) {
return 1;
} else {
return 0;
}
}
}
private void buildClusters() throws TasteException {
DataModel model = getDataModel();
int numUsers = model.getNumUsers();
if (numUsers == 0) {
topRecsByUserID = new FastByIDMap<List<RecommendedItem>>();
clustersByUserID = new FastByIDMap<FastIDSet>();
} else {
List<FastIDSet> clusters = Lists.newArrayList();
// Begin with a cluster for each user:
LongPrimitiveIterator it = model.getUserIDs();
while (it.hasNext()) {
FastIDSet newCluster = new FastIDSet();
newCluster.add(it.nextLong());
clusters.add(newCluster);
}
boolean done = false;
while (!done) {
done = mergeClosestClusters(numUsers, clusters, done);
}
topRecsByUserID = computeTopRecsPerUserID(clusters);
clustersByUserID = computeClustersPerUserID(clusters);
allClusters = clusters.toArray(new FastIDSet[clusters.size()]);
}
}
private boolean mergeClosestClusters(int numUsers, List<FastIDSet> clusters, boolean done) throws TasteException {
// We find a certain number of closest clusters...
List<ClusterClusterPair> queue = findClosestClusters(numUsers, clusters);
// The first one is definitely the closest pair in existence so we can cluster
// the two together, put it back into the set of clusters, and start again. Instead
// we assume everything else in our list of closest cluster pairs is still pretty good,
// and we cluster them too.
while (!queue.isEmpty()) {
if (!clusteringByThreshold && clusters.size() <= numClusters) {
done = true;
break;
}
ClusterClusterPair top = queue.remove(0);
if (clusteringByThreshold && top.getSimilarity() < clusteringThreshold) {
done = true;
break;
}
FastIDSet cluster1 = top.getCluster1();
FastIDSet cluster2 = top.getCluster2();
// Pull out current two clusters from clusters
Iterator<FastIDSet> clusterIterator = clusters.iterator();
boolean removed1 = false;
boolean removed2 = false;
while (clusterIterator.hasNext() && !(removed1 && removed2)) {
FastIDSet current = clusterIterator.next();
// Yes, use == here
if (!removed1 && cluster1 == current) {
clusterIterator.remove();
removed1 = true;
} else if (!removed2 && cluster2 == current) {
clusterIterator.remove();
removed2 = true;
}
}
// The only catch is if a cluster showed it twice in the list of best cluster pairs;
// have to remove the others. Pull out anything referencing these clusters from queue
for (Iterator<ClusterClusterPair> queueIterator = queue.iterator(); queueIterator.hasNext();) {
ClusterClusterPair pair = queueIterator.next();
FastIDSet pair1 = pair.getCluster1();
FastIDSet pair2 = pair.getCluster2();
if (pair1 == cluster1 || pair1 == cluster2 || pair2 == cluster1 || pair2 == cluster2) {
queueIterator.remove();
}
}
// Make new merged cluster
FastIDSet merged = new FastIDSet(cluster1.size() + cluster2.size());
merged.addAll(cluster1);
merged.addAll(cluster2);
// Compare against other clusters; update queue if needed
// That new pair we're just adding might be pretty close to something else, so
// catch that case here and put it back into our queue
for (FastIDSet cluster : clusters) {
double similarity = clusterSimilarity.getSimilarity(merged, cluster);
if (similarity > queue.get(queue.size() - 1).getSimilarity()) {
ListIterator<ClusterClusterPair> queueIterator = queue.listIterator();
while (queueIterator.hasNext()) {
if (similarity > queueIterator.next().getSimilarity()) {
queueIterator.previous();
break;
}
}
queueIterator.add(new ClusterClusterPair(merged, cluster, similarity));
}
}
// Finally add new cluster to list
clusters.add(merged);
}
return done;
}
private List<ClusterClusterPair> findClosestClusters(int numUsers,
List<FastIDSet> clusters) throws TasteException {
Queue<ClusterClusterPair> queue =
new PriorityQueue<ClusterClusterPair>(numUsers + 1, Collections.<ClusterClusterPair>reverseOrder());
int size = clusters.size();
for (int i = 0; i < size; i++) {
FastIDSet cluster1 = clusters.get(i);
for (int j = i + 1; j < size; j++) {
FastIDSet cluster2 = clusters.get(j);
double similarity = clusterSimilarity.getSimilarity(cluster1, cluster2);
if (!Double.isNaN(similarity)) {
if (queue.size() < numUsers) {
queue.add(new ClusterClusterPair(cluster1, cluster2, similarity));
} else if (similarity > queue.poll().getSimilarity()) {
queue.add(new ClusterClusterPair(cluster1, cluster2, similarity));
queue.poll();
}
}
}
}
List<ClusterClusterPair> result = Lists.newArrayList(queue);
Collections.sort(result);
return result;
}
private FastByIDMap<List<RecommendedItem>> computeTopRecsPerUserID(Iterable<FastIDSet> clusters)
throws TasteException {
FastByIDMap<List<RecommendedItem>> recsPerUser = new FastByIDMap<List<RecommendedItem>>();
for (FastIDSet cluster : clusters) {
List<RecommendedItem> recs = computeTopRecsForCluster(cluster);
LongPrimitiveIterator it = cluster.iterator();
while (it.hasNext()) {
recsPerUser.put(it.nextLong(), recs);
}
}
return recsPerUser;
}
private List<RecommendedItem> computeTopRecsForCluster(FastIDSet cluster) throws TasteException {
DataModel dataModel = getDataModel();
FastIDSet possibleItemIDs = new FastIDSet();
LongPrimitiveIterator it = cluster.iterator();
while (it.hasNext()) {
possibleItemIDs.addAll(dataModel.getItemIDsFromUser(it.nextLong()));
}
TopItems.Estimator<Long> estimator = new Estimator(cluster);
List<RecommendedItem> topItems = TopItems.getTopItems(NUM_CLUSTER_RECS,
possibleItemIDs.iterator(), null, estimator);
log.debug("Recommendations are: {}", topItems);
return Collections.unmodifiableList(topItems);
}
private static FastByIDMap<FastIDSet> computeClustersPerUserID(Collection<FastIDSet> clusters) {
FastByIDMap<FastIDSet> clustersPerUser = new FastByIDMap<FastIDSet>(clusters.size());
for (FastIDSet cluster : clusters) {
LongPrimitiveIterator it = cluster.iterator();
while (it.hasNext()) {
clustersPerUser.put(it.nextLong(), cluster);
}
}
return clustersPerUser;
}
@Override
public void refresh(Collection<Refreshable> alreadyRefreshed) {
refreshHelper.refresh(alreadyRefreshed);
}
@Override
public String toString() {
return "TreeClusteringRecommender2[clusterSimilarity:" + clusterSimilarity + ']';
}
private final class Estimator implements TopItems.Estimator<Long> {
private final FastIDSet cluster;
private Estimator(FastIDSet cluster) {
this.cluster = cluster;
}
@Override
public double estimate(Long itemID) throws TasteException {
DataModel dataModel = getDataModel();
RunningAverage average = new FullRunningAverage();
LongPrimitiveIterator it = cluster.iterator();
while (it.hasNext()) {
Float pref = dataModel.getPreferenceValue(it.nextLong(), itemID);
if (pref != null) {
average.addDatum(pref);
}
}
return average.getAverage();
}
}
}