net.librec.recommender.cf.ranking

Class LDARecommender

java.lang.Object

net.librec.recommender.AbstractRecommender

net.librec.recommender.ProbabilisticGraphicalRecommender

net.librec.recommender.cf.ranking.LDARecommender

All Implemented Interfaces:

Recommender

@ModelData(value={"isRanking","lda","userTopicProbs","topicItemProbs","trainMatrix"})
public class LDARecommender
extends ProbabilisticGraphicalRecommender

Latent Dirichlet Allocation for implicit feedback: Tom Griffiths, Gibbs sampling in the generative model of Latent Dirichlet Allocation, 2002.

Remarks: This implementation of LDA is for implicit feedback, where users are regarded as documents and items as words. To directly apply LDA to explicit ratings, Ian Porteous et al. (AAAI 2008, Section Bi-LDA) mentioned that, one way is to treat items as documents and ratings as words. We did not provide such an LDA implementation for explicit ratings. Instead, we provide recommender URP as an alternative LDA model for explicit ratings.

Author:

guoguibing and Keqiang Wang

Field Summary

Fields

Modifier and Type	Field and Description
protected DenseVector	alpha vector of hyperparameters for alpha and beta
protected DenseVector	beta vector of hyperparameters for alpha and beta
protected float	initAlpha Dirichlet hyper-parameters of user-topic distribution: typical value is 50/K
protected float	initBeta Dirichlet hyper-parameters of topic-item distribution, typical value is 0.01
protected int	numStats size of statistics
protected int	numTopics number of topics
protected java.util.List<java.lang.Integer>	topicAssignments topic assignment as list from the iterator of trainMatrix
protected DenseMatrix	topicItemNumbers entry[k, i]: number of tokens assigned to topic k, given item i.
protected DenseMatrix	topicItemProbs posterior probabilities of parameters
protected DenseMatrix	topicItemProbsSum cumulative statistics of theta, phi
protected DenseVector	topicTokenNumbers entry[k]: number of tokens assigned to topic t.
protected DenseVector	userTokenNumbers entry[u]: number of tokens rated by user u.
protected DenseMatrix	userTopicNumbers entry[u, k]: number of tokens assigned to topic k, given user u.
protected DenseMatrix	userTopicProbs posterior probabilities of parameters
protected DenseMatrix	userTopicProbsSum cumulative statistics of theta, phi

Fields inherited from class net.librec.recommender.ProbabilisticGraphicalRecommender

burnIn, numItems, numIterations, numUsers, sampleLag

Fields inherited from class net.librec.recommender.AbstractRecommender

conf, context, decay, earlyStop, globalMean, isBoldDriver, isRanking, itemMappingData, lastLoss, LOG, loss, maxRate, minRate, numRates, ratingScale, recommendedList, testMatrix, topN, trainMatrix, userMappingData, validMatrix, verbose

Constructor Summary

Constructors

Constructor and Description
LDARecommender()

Method Summary

Methods

Modifier and Type	Method and Description
protected void	eStep() parameters estimation: used in the training phase
protected void	estimateParams() estimate the model parameters
protected void	mStep() update the hyper-parameters
protected double	predict(int userIdx, int itemIdx) predict a specific rating for user userIdx on item itemIdx, note that the prediction is not bounded.
protected void	readoutParams() Add to the statistics the values of theta and phi for the current state.
protected void	setup() setup init member method

Methods inherited from class net.librec.recommender.ProbabilisticGraphicalRecommender

isConverged, trainModel

Methods inherited from class net.librec.recommender.AbstractRecommender

cleanup, evaluate, evaluateMap, getContext, getDataModel, getRecommendedList, loadModel, predict, recommend, recommend, recommendRank, recommendRating, saveModel, setContext

Methods inherited from class java.lang.Object

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

Field Detail

initAlpha

protected float initAlpha

Dirichlet hyper-parameters of user-topic distribution: typical value is 50/K

initBeta

protected float initBeta

Dirichlet hyper-parameters of topic-item distribution, typical value is 0.01

topicItemNumbers

protected DenseMatrix topicItemNumbers

entry[k, i]: number of tokens assigned to topic k, given item i.

userTopicNumbers

protected DenseMatrix userTopicNumbers

entry[u, k]: number of tokens assigned to topic k, given user u.

topicAssignments

protected java.util.List<java.lang.Integer> topicAssignments

topic assignment as list from the iterator of trainMatrix

userTokenNumbers

protected DenseVector userTokenNumbers

entry[u]: number of tokens rated by user u.

topicTokenNumbers

protected DenseVector topicTokenNumbers

entry[k]: number of tokens assigned to topic t.

numTopics

protected int numTopics

number of topics

alpha

protected DenseVector alpha

vector of hyperparameters for alpha and beta

beta

protected DenseVector beta

vector of hyperparameters for alpha and beta

userTopicProbsSum

protected DenseMatrix userTopicProbsSum

cumulative statistics of theta, phi

topicItemProbsSum

protected DenseMatrix topicItemProbsSum

cumulative statistics of theta, phi

userTopicProbs

protected DenseMatrix userTopicProbs

posterior probabilities of parameters

topicItemProbs

protected DenseMatrix topicItemProbs

posterior probabilities of parameters

numStats

protected int numStats

size of statistics

Constructor Detail

LDARecommender

public LDARecommender()

Method Detail

setup

protected void setup()
              throws LibrecException

setup init member method

Overrides:

setup in class ProbabilisticGraphicalRecommender

Throws:

LibrecException - if error occurs

eStep

protected void eStep()

Description copied from class: ProbabilisticGraphicalRecommender

parameters estimation: used in the training phase

Specified by:

eStep in class ProbabilisticGraphicalRecommender

mStep

protected void mStep()

Description copied from class: ProbabilisticGraphicalRecommender

update the hyper-parameters

Specified by:

mStep in class ProbabilisticGraphicalRecommender

readoutParams

protected void readoutParams()

Add to the statistics the values of theta and phi for the current state.

Overrides:

readoutParams in class ProbabilisticGraphicalRecommender

estimateParams

protected void estimateParams()

Description copied from class: ProbabilisticGraphicalRecommender

estimate the model parameters

Overrides:

estimateParams in class ProbabilisticGraphicalRecommender

predict

protected double predict(int userIdx,
             int itemIdx)
                  throws LibrecException

Description copied from class: AbstractRecommender

predict a specific rating for user userIdx on item itemIdx, note that the prediction is not bounded. It is useful for building models with no need to bound predictions.

Specified by:

predict in class AbstractRecommender

Parameters:

userIdx - user index

itemIdx - item index

Returns:

predictive rating for user userIdx on item itemIdx without bound

Throws:

LibrecException - if error occurs during predicting