Models Taxonomy

WarpRec ships with 70+ built-in algorithms spanning 6 model families. All models can run locally or at cluster scale via Ray.

Family	Model	Description
Unpersonalized	Pop	Recommends the most popular items overall.
	Random	Recommends items uniformly at random (lower-bound baseline).
	ProxyRecommender	Evaluates precomputed recommendation lists from an external file.
Content-Based	VSM	Vector Space Model using TF-IDF and cosine similarity on item profiles.
CF / Autoencoder	EASE	Closed-form linear autoencoder via ridge regression for item similarity.
	ELSA	Scalable EASE approximation using sparse low-rank SGD decomposition.
	CDAE	Denoising autoencoder with user-specific latent vectors.
	MacridVAE	Disentangled VAE modeling macro user intentions via concept routing.
	MultiDAE	Multinomial denoising autoencoder for implicit feedback.
	MultiVAE	Variational autoencoder with reparameterization for implicit feedback.
	SANSA	Sparse approximate non-symmetric autoencoder via LDL^T decomposition.
CF / Graph-Based	DGCF	Disentangled graph CF with iterative factor routing.
	EGCF	Embedding-less graph CF using BPR + InfoNCE contrastive learning.
	ESIGCF	Extremely simplified intent-enhanced graph CF with JoGCN.
	GCMC	Graph convolutional matrix completion for explicit feedback.
	LightCCF	Contrastive CF with Neighborhood Aggregation loss (MF or GCN encoder).
	LightGCL	Graph contrastive learning using SVD for global view augmentation.
	LightGCN	Simplified GCN with linear propagation (no feature transforms).
	LightGCN++	LightGCN with asymmetric normalization and residual connections.
	LightGODE	Training-free graph convolution; applies ODE solver at inference.
	MACRGCN	Counterfactual reasoning on LightGCN backbone to eliminate popularity bias.
	MixRec	Dual individual/collective mixing with contrastive learning.
	NGCF	Neural graph CF with higher-order connectivity propagation.
	PAAC	Popularity-aware alignment and contrast for mitigating popularity bias.
	PopDCL	Popularity-aware debiased contrastive loss on LightGCN backbone.
	RecDCL	Dual contrastive learning combining feature-wise and topology-wise CL.
	RP3Beta	Biased random walk of length 3 on the user-item bipartite graph.
	SGCL	Supervised graph contrastive learning without negative sampling.
	SGL	Self-supervised graph learning with structure augmentation (ED/ND/RW).
	SimGCL	Contrastive learning via uniform noise perturbation without graph augmentation.
	SimRec	Graph-less CF distilling GCN teacher into MLP student via contrastive KD.
	UltraGCN	Infinite-layer GCN approximation via constraint losses (no message passing).
	XSimGCL	Graph contrastive learning with uniform noise perturbation.
CF / KNN	ItemKNN	Item-based collaborative KNN using configurable similarity metrics.
	ItemKNN-TD	Item-based KNN with temporal decay weighting on interactions.
	UserKNN	User-based collaborative KNN from historical interactions.
	UserKNN-TD	User-based KNN with temporal decay weighting on interactions.
CF / Latent Factor	ADMMSlim	Sparse item similarity matrix optimized via ADMM.
	BPR	Bayesian Personalized Ranking for pairwise implicit feedback.
	FISM	Factored item similarity with weighted-average user representations.
	iALS	Implicit ALS revisited for top-N recommendation with all-pairs weighting.
	iALS2008	Original confidence-weighted implicit ALS for implicit-feedback matrix factorization.
	MACRMF	Counterfactual MF eliminating popularity bias via causal inference.
	Slim	Sparse linear method with L1/L2 (ElasticNet) regularization.
CF / Neural	ConvNCF	CNN on user-item embedding outer product for structured interaction patterns.
	NeuMF	Hybrid neural CF combining GMF and MLP branches.
Context-Aware	AFM	Attentional Factorization Machine with attention-weighted feature interactions.
	DCN	Deep & Cross Network for explicit bounded-degree feature crossing.
	DCNv2	Improved DCN with Mixture-of-Experts and low-rank cross layers.
	DeepFM	Parallel FM + DNN for low-order and high-order feature interactions.
	FM	Factorization Machine modeling second-order feature interactions.
	NFM	Neural FM with Bi-Interaction pooling layer followed by MLP.
	WideAndDeep	Joint wide (linear) + deep (DNN) model for memorization and generalization.
	xDeepFM	Compressed Interaction Network (CIN) for vector-wise explicit interactions.
Sequential / KNN	STAN	Sequence- and time-aware neighborhood model for session-based recommendation.
Sequential / CNN	Caser	Convolutional sequence embedding with horizontal and vertical filters.
Sequential / Markov	FOSSIL	First-order Markov chain fused with factored item similarity.
Sequential / RNN	GRU4Rec	Session-based GRU for short-term preference modeling.
	NARM	Hybrid GRU encoder with global + local attention mechanisms.
Sequential / Transformer	BERT4Rec	Bidirectional Transformer with masked item prediction (cloze task).
	BSARec	Bandlimited self-attention combining frequency filtering with Transformer attention.
	CL4SRec	SASRec-style encoder with sequence augmentations and contrastive learning.
	CORE	Consistent Representation Encoder unifying encoding/decoding spaces.
	DuoRec	Dual contrastive sequential recommendation with unsupervised and supervised CL.
	eSASRec	Enhanced SASRec variant with LiGR blocks and sampled-softmax support.
	gSASRec	General self-attention with group-wise binary cross-entropy loss.
	LightSANs	Low-rank decomposed self-attention with decoupled position encoding.
	LinRec	Linear attention mechanism (O(N)) for efficient long-sequence modeling.
	SASRec	Self-attentive Transformer for sequential recommendation.
Hybrid / Autoencoder	AddEASE	EASE extension solving two linear problems to incorporate side information.
	CEASE	Closed-form extended EASE via augmented interaction matrix with side info.
Hybrid / KNN	AttributeItemKNN	Item-based KNN using content features for similarity.
	AttributeUserKNN	User-based KNN using content-derived user profiles.