Recommender

Similarity

The similarity matrix is used to construct the distance between user and user or item and item in the data set. The example is shown as follows.

rec.similarity.class=cos
rec.recommender.similarities=user

rec.similarity.class is the distance measure function. rec.recommender.similarities is the distance of a specified user-user pair or a item-item pair. For the Social recommendation algorithms, users can also specify the distance of a social-social pair.

In the Java program, the corresponding sample program is shown as follows.

conf.set("rec.recommender.similarities","user");
RecommenderSimilarity similarity = new CosineSimilarity();
similarity.buildSimilarityMatrix(dataModel);

After generating the RecommenderSimilarity object, users need to call the buildSimilarityMatrix method to build the similarity matrix.

The table below is the distance measure functions and corresponding abbreviations in LibRec.

Similarity	shortname
BinaryCosineSimilarity	bcos
CosineSimilarity	cos
CPCSimilarity	cpc
MSESimilarity	msesim
MSDSimilarity	msd
PCCSimilarity	pcc
KRCCSimilarity	krcc
DiceCoefficientSimilarity	dice
JaccardSimilarity	jaccard
ExJaccardSimilarity	exjaccard

Algorithms

When users use the configuration and command line to run programs, the recommendation algorithm is specified by rec.recommender.class. The configuration is shown as follows.

rec.recommender.class=shortname #e.g. aobpr

Please refer to [Algorithm list.md](./Algorithm list) for the abbreviations of different algorithms.

In the Java implementation, after making instances of the Configuration object, the DataModel object, and the Similarity matrix object, these three instances are passed in as constructor parameters to generate the RecommenderContext object. Users can make the corresponding instance of the recommendation algorithm, that is to say, no need to set the rec.recommedner.class configuration. The example code is shown as follows.

RecommenderContext context = new RecommenderContext(conf, dataModel, similarity);

conf.set("rec.neighbors.knn.number","50");
conf.set("rec.recommender.isranking=false");

Recommender recommender = new UserKNNRecommender();
recommender.train(context);

Each recommendation algorithm has its own configurations according to different algorithm categories. Current recommendation algorithms implemented in LibRec are matrix-decomposition-based, factor-factorization-based, probability-graph-model-based, and tensor-based. The following part gives different configurations based on different algorithms. In general, taking the matrix factorization for example, making recommendations by inheriting the interface of matrix factorization algorithms not only needs to set the configuration of matrix factorization, but also needs to set other configurations. For example, the BPMF also needs to configure rec.recommender.user.mu and other configurations.

The implementation of the recommendation algorithms based on the above algorithms, which only needs to inherit the corresponding abstract classes. Different algorithms are listed in the following table. All the recommendation algorithm configurations are listed in Algorithm List.

AbstractRecommender

# if ranking
rec.recommender.isranking=true
rec.recommender.ranking.topn=10

The algorithms that directly implement the AbstractRecommender abstract class are shown as follows.

directory path	short name	algorithm
poi	usg	USGRecommender

MatrixRecommender

directory path	short name	algorithm
baseline	constantguess	ConstantGuessRecommender
baseline	globalaverage	GlobalAverageRecommender
baseline	itemaverage	ItemAverageRecommender
baseline	mostpopular	MostPopularRecommender
baseline	randomguess	RandomGuessRecommender
baseline	useraverage	UserAverageRecommender
cf	itemknn	ItemKNNRecommender
cf	userknn	UserKNNRecommender
cf.rating	rbm	RBMRecommender
ext	associationrule	AssociationRuleRecommender
ext	bipolarslopeone	BipolarSlopeOneRecommender
ext	external	ExternalRecommender
ext	personalitydiagnosis	PersonalityDiagnosisRecommender
ext	slopeone	SlopeOneRecommender
hybrid	hybrid	HybridRecommender
nn.ranking	cdae	CDAERecommender
nn.rating	autorec	AutoRecRecommender

Matrix Probabilistic Graphical Recommender

rec.iterator.maximum=1000
rec.pgm.burn-in=100
rec.pgm.samplelag=100

The algorithms that directly inherit the MatrixProbabilisticGraphicalRecommender class are shown as follows.

directory path	short name	algorithm
baseline	itemcluster	ItemClusterRecommender
baseline	usercluster	UserClusterRecommender
cf	bhfree	BHFreeRecommender
cf	bucm	BUCMRecommender
cf.ranking	aspectmodelranking	AspectModelRecommender
cf.ranking	itembigram	ItemBigramRecommender
cf.ranking	lda	LDARecommender
cf.ranking	plsa	PLSARecommender
cf.rating	aspectmodelrating	AspectModelRecommender
cf.rating	gplsa	GPLSARecommender
cf.rating	ldcc	LDCCRecommender
cf.rating	urp	URPRecommender

Matrix Factorization Recommender

rec.iterator.maximum=100
rec.iterator.learningrate=0.01
rec.iterator.learningrate.maximum=1000
rec.user.regularization=0.01
rec.item.regularization=0.01
rec.factor.number=10
rec.learningrate.bolddriver=false
rec.learningrate.decay=1.0

The algorithms that directly or indirectly inherit the Matrix Factorization Recommender class are shown as follows.

directory path	short name	algorithm
cf.rating	asvdpp	ASVDPlusPlusRecommender
cf.rating	svdpp	SVDPlusPlusRecommender
content	efm	EFMRecommender
content	hft	HFTRecommender
context.rating	timesvd	TimeSVDRecommender
cf.ranking	aobpr	AoBPRRecommender
cf.ranking	bnppf	BNPPFRecommeder
cf.ranking	bpoissmf	BPoissMFRecommender
cf.ranking	bpr	BPRRecommender
cf.ranking	climf	CLIMFRecommender
cf.ranking	cofiset	CoFiSetRecommender
cf.ranking	eals	EALSRecommender
cf.ranking	fismauc	FISMaucRecommender
cf.ranking	fismrmse	FISMrmseRecommender
cf.ranking	gbpr	GBPRRecommender
cf.ranking	listwisemf	ListwiseMFRecommender
cf.ranking	nmfitemitem	NMFItemItemRecommender
cf.ranking	pnmf	PNMFRecommender
cf.ranking	rankals	RankALSRecommender
cf.ranking	rankpmf	RankPMFRecommender
cf.ranking	ranksgd	RankSGDRecommender
cf.ranking	slim	SLIMRecommender
cf.ranking	wbpr	WBPRRecommender
cf.ranking	wrmf	WRMFRecommender
cf.rating	biasedmf	BiasedMFRecommender
cf.rating	bnpoissmf	BNPoissMFRecommender
cf.rating	bpmf	BPMFRecommender
cf.rating	irrg	IRRGRecommender
cf.rating	llorma	LLORMARecommender
cf.rating	mfals	MFALSRecommender
cf.rating	nmf	NMFRecommender
cf.rating	pmf	PMFRecommender
cf.rating	rbm	RBMRecommender
cf.rating	remf	ReMFRecommender
cf.rating	rfrec	RFRecRecommender
poi	rankgeofm	RankGeoFMRecommender

Factorization Machine Recommender

rec.recommender.maxrate=12.0
rec.recommender.minrate=0.0

rec.factor.number=10

rec.fm.regw0=0.01
reg.fm.regW=0.01
reg.fm.regF=10

The algorithms that directly inherit the Factorization Machine Recommender class are shown as follows.

directory path	short name	algorithm
cf.rating	ffm	FFMRecommender
cf.rating	fmals	FMALSRecommender
cf.rating	fmftrl	FMFTRLRecommender
cf.rating	fmsgd	FMSGDRecommender
context.ranking	dlambdafm	DLambdaFMRecommender

Social recommender

rec.iterator.maximum=100
rec.iterator.learningrate=0.01
rec.iterator.learningrate.maximum=1000
rec.user.regularization=0.01
rec.item.regularization=0.01
rec.factor.number=10
rec.learningrate.bolddriver=false
rec.learningrate.decay=1.0

rec.social.regularization=0.01

The algorithms that directly inherit the Social Recommender class are shown as follows.

directory path	short name	algorithm
context.rating	trustmf	TrustMFRecommender
context.ranking	sbpr	SBPRRecommender
context.rating	rste	RSTERecommender
context.rating	socialmf	SocialMFRecommender
context.rating	sorec	SoRecRecommender
context.rating	soreg	SoRegRecommender
context.rating	trustsvd	TrustSVDRecommender

TensorRecommender

rec.recommender.verbose=true
rec.iterator.learningrate=0.01
rec.iterator.learningrate.maximum=1000
rec.factor.number=10
rec.tensor.regularization=0.01

The algorithms that directly inherit the Tensor Recommender class are shown as follows.

directory path	short name	algorithm
content	convmf	ConvMFRecommender
content	efm	EFMRecommender
content	hft	HFTRecommender
content	tfidf	TFIDFRecommender
content	topicmfat	TopicMFATRecommender
content	topicmfmt	TopicMFMTRecommender
context.rating	cptf	CPTFRecommender

Implement your own algorithm

Implementing users' own algorithms in LibRec needs to inherit the corresponding abstract class according to the algorithm's category. And users need to implement the abstract methods, or re-write the methods in the abstract class. In order to inherit the AbstractRecommender class, for example, the implementation of an algorithm is shown as follows.

1. Re-write the setup method (optional)
    The main task of the setup method is to initialize the member variables. For example, reading arguments from the configuration file. If the algorithm does not need to set extra configurations, the program does not need to re-write the methods.
    It should be noted that, in the rewrite of the setup method, users need to call the original abstract class in the setup method. That is, to put super.setup() in the first line of the implementation, to ensure that the basic parameters of the algorithm is initialized.

2. Implement the trainModel Method
    The trainModel method accomplish the task of training the model. For example, using gradient descent to optimize the cost function of the model.

3. Implement the Predict Method
    The task of the Predict method is to use the trained model to make predictions.
    For example, for rating prediction algorithms, the predict method is used to predict each rating in the test set. Particularly, for a given userIndex and itemIndex, using the model to predict the rating score of userIndex-itemIndex pair.

Please refer to the exact implementation in the recommender directory.