Model Documentation

Welcome to the CoreRec model documentation. This section provides detailed information about all available recommendation models.

Model Tiers

CoreRec organizes its models into two tiers:

Production Models (Tested & Stable)

These 14 models are fully tested, CI-enforced, and recommended for production use. They implement the full BaseRecommender interface (fit, predict, recommend, save, load) and are covered by automated tests on every commit.

Model	Type	Description
DCN	Deep Learning	Deep & Cross Network for explicit feature interactions
DeepFM	Deep Learning	Factorization Machines + Deep Neural Networks
GNNRec	Graph	Graph Neural Networks for recommendations
MIND	Multi-Interest	Multi-Interest Network with dynamic routing
NASRec	AutoML	Neural Architecture Search for recommendations
SASRec	Sequential	Self-Attentive Sequential Recommendation
TwoTower	Retrieval	Dual-encoder for large-scale candidate retrieval
BERT4Rec	Sequential	Bidirectional self-attention for sequence modeling
SAR	Collaborative	Smart Adaptive Recommendations (co-occurrence)
NCF	Collaborative	Neural Collaborative Filtering (MLP + GMF)
FAST	Collaborative	Fast Approximate Similarity-based filtering
FASTRecommender	Collaborative	Extended FAST with additional features
LightGCN	Graph	Simplified graph convolution for CF
TFIDFRecommender	Content-Based	TF-IDF content similarity recommendations

Production Guarantee

All production models pass automated tests on every push/PR. If a production model test fails, the commit is blocked from merging.

Sandbox Models (Experimental)

CoreRec also includes ~50 sandbox models in corerec/sandbox/. These are experimental implementations of well-known recommendation algorithms intended for research, learning, and prototyping.

Sandbox Warning

Sandbox models are not production-tested. They may have incomplete implementations, missing methods, or untested edge cases. Use them for exploration and research — not for production deployments without thorough validation.

Sandbox models cover a wide range of architectures:

Deep Learning — AFM, AutoInt, BST, BiVAE, Caser, DeepCrossing, DeepRec, DIEN, DiFM, DIN, DLRM, ENSFM, ESCM2, ESMM, FGCNN, FFM, FiBiNet, FLEN, FM, GAN-Rec, GateNet, GRU-CF, NFM, NextItNet, Wide&Deep, YouTubeDNN, PNN, MMoE, PLE, Monolith, TDM, AutoFI

Matrix Factorization — ALS, SVD, A2SVD, MF-Base, FM-Base, UserBased CF, Matrix Factorization

Graph-Based — GeoIMC, LightGCN-Base, GNN-Base

Sequential — RBM, RLRMC, SLi-Rec, SUM

Bayesian — BPR, BPR-MF, VMF

Content-Based — MIND-Content

Model Categories

Deep Learning Models

Neural network-based recommendation models that learn complex patterns from user-item interactions.

Production: DCN, DeepFM, GNNRec, MIND, NASRec, SASRec, TwoTower, BERT4Rec Sandbox: AFM, AutoInt, BST, BiVAE, Caser, and 25+ more

Matrix Factorization

Classic collaborative filtering approaches using matrix decomposition.

Production: SAR, NCF, FAST, FASTRecommender Sandbox: ALS, SVD, A2SVD, MF-Base, and more

Graph-Based Models

Models that leverage graph structure and relationships.

Production: GNNRec, LightGCN Sandbox: GeoIMC, LightGCN-Base, GNN-Base

Sequential Models

Time-aware models that consider user behavior sequences.

Production: SASRec, BERT4Rec Sandbox: RBM, RLRMC, SLi-Rec, SUM

Bayesian Models

Probabilistic approaches to recommendation.

Sandbox only: BPR, BPR-MF, VMF

Content-Based Models

Feature-based recommendations using item/user attributes.

Production: TFIDFRecommender Sandbox: MIND-Content

Model Selection Guide

When to Use Deep Learning Models

• Large datasets with complex patterns

• Rich feature sets available

• Need to capture non-linear relationships

When to Use Matrix Factorization

• Sparse user-item matrices

• Need interpretable recommendations

• Fast training and inference required

When to Use Graph-Based Models

• Rich relationship data available

• Social networks or knowledge graphs

• Need to leverage item/item or user/user connections

When to Use Sequential Models

• Temporal patterns important

• User behavior sequences available

• Need next-item prediction

When to Use Bayesian Models

• Uncertainty quantification needed

• Probabilistic interpretation desired

• Cold start problems

Getting Started

1. Start with production models — they’re tested and guaranteed to work

2. Browse the Complete Model Index for imports and tutorial links

3. Read the specific model tutorial for code examples

4. Only explore sandbox models if you need a specific algorithm not covered by production models

5. If using a sandbox model in production, validate thoroughly and contribute tests back

Tutorials

For hands-on examples, see the Tutorials section.