org.grouplens.lenskit.mf.funksvd

Class FunkSVDModelBuilder

java.lang.Object

org.grouplens.lenskit.mf.funksvd.FunkSVDModelBuilder

All Implemented Interfaces:

javax.inject.Provider<FunkSVDModel>

public class FunkSVDModelBuilder
extends Object
implements javax.inject.Provider<FunkSVDModel>

SVD recommender builder using gradient descent (Funk SVD).

This recommender builder constructs an SVD-based recommender using gradient descent, as pioneered by Simon Funk. It also incorporates the regularizations Funk did. These are documented in Netflix Update: Try This at Home. This implementation is based in part on Timely Development's sample code.

Field Summary

Fields

Modifier and Type	Field and Description
protected int	featureCount
protected double	initialValue
protected FunkSVDUpdateRule	rule
protected PreferenceSnapshot	snapshot

Constructor Summary

Constructors

Constructor and Description
FunkSVDModelBuilder(PreferenceSnapshot snapshot, FunkSVDUpdateRule rule, int featureCount, double initVal)

Method Summary

Modifier and Type	Method and Description
protected double	doFeatureIteration(TrainingEstimator estimates, Collection<IndexedPreference> ratings, mikera.vectorz.Vector userFeatureVector, mikera.vectorz.Vector itemFeatureVector, double trail) Do a single feature iteration.
FunkSVDModel	get()
protected void	summarizeFeature(mikera.vectorz.AVector ufv, mikera.vectorz.AVector ifv, FeatureInfo.Builder fib) Add a feature's summary to the feature info builder.
protected void	trainFeature(int feature, TrainingEstimator estimates, mikera.vectorz.Vector userFeatureVector, mikera.vectorz.Vector itemFeatureVector, FeatureInfo.Builder fib) Train a feature using a collection of ratings.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

featureCount

protected final int featureCount

snapshot

protected final PreferenceSnapshot snapshot

initialValue

protected final double initialValue

rule

protected final FunkSVDUpdateRule rule

Constructor Detail

FunkSVDModelBuilder

@Inject
public FunkSVDModelBuilder(@Nonnull
                                   PreferenceSnapshot snapshot,
                                   @Nonnull
                                   FunkSVDUpdateRule rule,
                                   @FeatureCount
                                   int featureCount,
                                   @InitialFeatureValue
                                   double initVal)

Method Detail

get

public FunkSVDModel get()

Specified by:

get in interface javax.inject.Provider<FunkSVDModel>

trainFeature

protected void trainFeature(int feature,
                            TrainingEstimator estimates,
                            mikera.vectorz.Vector userFeatureVector,
                            mikera.vectorz.Vector itemFeatureVector,
                            FeatureInfo.Builder fib)

Train a feature using a collection of ratings. This method iteratively calls doFeatureIteration(TrainingEstimator, Collection, Vector, Vector, double) to train the feature. It can be overridden to customize the feature training strategy.

We use the estimator to maintain the estimate up through a particular feature value, rather than recomputing the entire kernel value every time. This hopefully speeds up training. It means that we always tell the updater we are training feature 0, but use a subvector that starts with the current feature.

Parameters:

feature - The number of the current feature.

estimates - The current estimator. This method is not expected to update the estimator.

userFeatureVector - The user feature values. This has been initialized to the initial value, and may be reused between features.

itemFeatureVector - The item feature values. This has been initialized to the initial value, and may be reused between features.

fib - The feature info builder. This method is only expected to add information about its training rounds to the builder; the caller takes care of feature number and summary data.

See Also:

doFeatureIteration(TrainingEstimator, Collection, Vector, Vector, double), summarizeFeature(AVector, AVector, FeatureInfo.Builder)

doFeatureIteration

protected double doFeatureIteration(TrainingEstimator estimates,
                                    Collection<IndexedPreference> ratings,
                                    mikera.vectorz.Vector userFeatureVector,
                                    mikera.vectorz.Vector itemFeatureVector,
                                    double trail)

Do a single feature iteration.

Parameters:

estimates - The estimates.

ratings - The ratings to train on.

userFeatureVector - The user column vector for the current feature.

itemFeatureVector - The item column vector for the current feature.

trail - The sum of the remaining user-item-feature values.

Returns:

The RMSE of the feature iteration.

summarizeFeature

protected void summarizeFeature(mikera.vectorz.AVector ufv,
                                mikera.vectorz.AVector ifv,
                                FeatureInfo.Builder fib)

Add a feature's summary to the feature info builder.

Parameters:

ufv - The user values.

ifv - The item values.

fib - The feature info builder.