EdoardoBotta/RQ-VAE-Recommender

[Pytorch] Generative retrieval model using semantic IDs from "Recommender Systems with Generative Retrieval"

What it solves

It addresses the complexity of implementing generative recommendation systems, where items are treated as tokens rather than just IDs. Instead of just providing a neural layer, it offers a complete experimental pipeline to handle semantic-ID tokenization, dataset preparation, identifier collisions, and the training of sequence models for item retrieval.

How it works

The project implements a two-stage workflow:

  1. Semantic-ID Tokenization: A Residual-Quantized Variational Autoencoder (RQ-VAE) maps catalog items into short tuples of discrete semantic IDs.
  2. Generative Retrieval: A Transformer (specifically a T5 encoder-decoder) takes a user's history of these semantic IDs and generates the IDs of the most likely next items.

Who it’s for

It is designed for recommender-system researchers and practitioners focusing on semantic tokenization, vector quantization, cold-item representations, and generative retrieval.

Highlights

  • Flexible Quantization: Supports K-means initialization and multiple gradient estimators including Gumbel-Softmax, straight-through, and the rotation-trick.
  • Integrated Pipeline: Includes automatic preprocessing for Amazon Reviews and MovieLens datasets.
  • Collision Handling: Features semantic-ID collision disambiguation and corpus-prefix-constrained generation.
  • Research Ready: Provides a T5-based retrieval model with built-in hit-rate and NDCG evaluation metrics.

Related

  • Project
    AkaliKong/MiniOneRec

    An open-source generative recommendation framework that converts items into semantic IDs (SIDs) and uses SFT and reinforcement learning to predict the next item a user will consume.

  • Project
    lucidrains/vector-quantize-pytorch

    A PyTorch library implementing various vector quantization techniques, including VQ, Residual VQ, and FSQ, to map continuous data into discrete representations for generative modeling.

  • Project
    inseq-team/inseq

    A PyTorch-based toolkit for post-hoc interpretability of sequence generation models, providing various feature attribution methods to explain model outputs.

  • Project
    alibaba/TorchEasyRec

    A PyTorch-based recommendation system framework that provides pre-implemented state-of-the-art models and scalable infrastructure for production-ready deep learning recommendations.

  • Project
    jacobgil/pytorch-grad-cam

    A PyTorch library for explainable AI in computer vision that provides a wide range of pixel attribution methods to visualize and diagnose model predictions.