Model Library Overview

Torch-RecHub provides a rich recommendation model library that covers ranking, matching, multi-task learning, and generative recommendation. All models are implemented in PyTorch and are easy to use and extend.

Model Library Structure

The model library is organized by recommendation stage and task type:

1. Ranking Models: Used during fine ranking to predict click-through rates or user preference scores for items
2. Matching Models: Used during candidate generation to retrieve candidates from a large item collection
3. Multi-Task Models: Jointly optimize multiple related tasks to improve generalization
4. Generative Recommendation: Uses generative models to produce personalized recommendations

Model Selection Guide

Choosing a Ranking Model

Model	Suitable Scenario	Characteristics
WideDeep	Basic ranking tasks	Combines linear and deep models to balance memorization and generalization
DeepFM	Scenarios where feature interactions matter	Captures both low-order and high-order feature interactions
DCN/DCNv2	Explicit feature-crossing scenarios	Explicitly learns high-order feature crosses with high computational efficiency
DIN	Scenarios with dynamically changing user interests	Uses attention to capture user interests
DIEN	Long-sequence interest modeling	Models the dynamic evolution of user interests
BST	Scenarios where sequence features matter	Uses a Transformer to model sequence features
AutoInt	Automatic feature-interaction learning	Automatically learns feature-interaction patterns

Choosing a Matching Model

Model	Suitable Scenario	Characteristics
DSSM	Text-matching scenarios	Uses a two-tower architecture to map users and items into the same vector space
YoutubeDNN	Large-scale recommendation	Deep matching based on user behavior sequences
MIND	Multi-interest recommendation	Learns multiple interest representations for each user
GRU4Rec/SASRec	Sequential recommendation	Models a user's recent behavior sequence
ComirecDR/ComirecSA	Controllable multi-interest recommendation	Allows control over the number of generated interests

Choosing a Multi-Task Model

Model	Suitable Scenario	Characteristics
SharedBottom	Scenarios with strongly related tasks	All tasks share the bottom network
MMOE	Scenarios with substantial task conflict	Uses a multi-gate mixture of experts so each task learns a different expert combination
PLE	Complex multi-task scenarios	Uses progressive layered extraction to alleviate negative transfer
ESMM	Scenarios with sample-selection bias	Uses entire-space modeling to address sample-selection bias
AITM	Scenarios with dependencies between tasks	Uses adaptive information transfer to learn task dependencies

Choosing a Generative Recommendation Model

Model	Suitable Scenario	Characteristics
HSTU	Next-item sequential recommendation	Hierarchical sequential transduction units with positional and temporal biases
HLLM	Sequential recommendation with precomputed LLM item embeddings	Freezes the item semantic table and trains a user-sequence Transformer
RQ-VAE	Item semantic-ID quantization	Compresses continuous item embeddings into multi-level codebook IDs
TIGER	Semantic-ID generative retrieval	Uses T5 to generate a valid semantic ID for the next item

Model Documentation

Ranking Models

Detailed descriptions of ranking-model principles, usage, and parameters.

View Ranking Model Documentation

Matching Models

Detailed descriptions of matching-model principles, usage, and parameters.

View Matching Model Documentation

Multi-Task Models

Detailed descriptions of multi-task-model principles, usage, and parameters.

View Multi-Task Model Documentation

Generative Recommendation Models

Detailed descriptions of generative recommendation-model principles, usage, and parameters.

View Generative Recommendation Model Documentation

Usage Example

# Ranking-model example
from torch_rechub.models.ranking import DeepFM
from torch_rechub.trainers import CTRTrainer

# Create the model
model = DeepFM(deep_features=deep_features, fm_features=fm_features, mlp_params={"dims": [256, 128], "dropout": 0.2})

# Create the trainer
trainer = CTRTrainer(model, optimizer_params={"lr": 0.001}, device="cpu")

# Train the model
trainer.fit(train_dataloader, val_dataloader)

# Matching-model example
from torch_rechub.models.matching import DSSM
from torch_rechub.trainers import MatchTrainer

# Create the model
model = DSSM(user_features=user_features, item_features=item_features, temperature=1.0,
             user_params={"dims": [256, 128, 64]}, item_params={"dims": [256, 128, 64]})

# Create the trainer
trainer = MatchTrainer(model, mode=0, device="cpu")

# Train the model
trainer.fit(train_dataloader)

Contributing a New Model

If you would like to contribute a new model, see the Contributing Guide and follow the project's coding standards and documentation requirements.