recommenders-team/recommenders

Best Practices on Recommendation Systems

What it solves

Building a recommendation system from scratch is hard: you need to try many algorithms, clean data into the right formats, measure quality with proper metrics, tune hyperparameters, and finally deploy the model. This project packages that whole workflow into one place, so researchers and developers can prototype and ship recommender systems faster.

How it works

It provides a collection of ready-to-use recommendation algorithms, from classic matrix factorization like ALS and SVD to deep learning models like NCF, LightGCN, and transformer-based sequential recommenders. Alongside the models, it includes utilities that handle common chores: loading datasets in algorithm-specific formats, splitting data into training and test sets, and computing offline evaluation metrics such as MAP, nDCG, precision, recall, RMSE, and MAE. The repository also contains Jupyter notebooks that walk through each of the five key tasks: data preparation, model building, evaluation, model selection and optimization, and operationalization (with Azure examples). A benchmark notebook shows how to compare different algorithms side by side on a standard dataset.

Who it’s for

Researchers, data scientists, and developers who want to experiment with or productionize recommendation systems, whether they are new to the field or experienced. It is also useful for students and enthusiasts who want to study how different algorithms behave.

Highlights

  • Covers more than 30 algorithms, including collaborative filtering, content-based filtering, and sequential/attention-based methods
  • Provides utility functions for data handling, splitting, and evaluation that work across many algorithms
  • Includes detailed Jupyter notebooks with both quick-start examples and deep dives explaining the math and implementation
  • Offers a benchmark notebook to compare algorithm performance on the same data
  • Supports multiple environments: CPU, GPU, and PySpark
  • Many algorithms are contributed by Microsoft, including SAR, xDeepFM, and DKN
  • Operationalization examples show how to deploy models in production on Azure

Related

  • Project
    MTSWebServices/RecTools

    A Python library that streamlines the creation of recommender systems by providing a unified interface for various models, from matrix factorization to advanced Transformers.

  • Project
    sb-ai-lab/RePlay

    An advanced framework for the development and evaluation of recommendation systems, covering the full pipeline from data preprocessing to production deployment.

  • Project
    NVIDIA/recsys-examples

    A collection of optimized recommender models and GPU libraries for large-scale ranking, retrieval, and generative recommendation, featuring agent-driven PyTorch optimization.

  • Project
    Puiching-Memory/TAAC_2026

    An experimental workspace for large-scale recommendation systems that unifies sequence modeling and feature interaction for conversion rate prediction.

  • Project
    meta-recsys/generative-recommenders

    Generative Recommenders is a research‑grade code base that reframes recommendation as a generative modeling task. It introduces the HSTU/HSTU‑Ultra attention layers and the M‑FALCON training algorithm, enabling trillion‑parameter sequential models that run 2‑3× faster (and up to 10×‑1000× overall) on modern GPUs. The repo provides scripts to reproduce state‑of‑the‑art results on MovieLens and Amazon datasets, tools for synthetic 3‑billion‑rating data, and FP8‑quantized kernels for production‑level inference. Apache 2.0 licensed and tested on Ubuntu 22.04, CUDA 12.4, Python 3.10.