Recommender Systems
Branch: master
Clone or download
Pull request Compare This branch is 11 commits ahead, 199 commits behind Microsoft:master.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


This repository provides examples and best practices for building recommendation systems, provided as Jupyter notebooks. The examples detail our learnings on five key tasks:

  • Prepare Data: Preparing and loading data for each recommender algorithm
  • Model: Building models using various recommender algorithms such as Alternating Least Squares (ALS), Singular Value Decomposition (SVD), etc.
  • Evaluate: Evaluating algorithms with offline metrics
  • Model Select and Optimize: Tuning and optimizing hyperparameteres for recommender models
  • Operationalize: Operationalizing models in a production environment on Azure

Several utilities are provided in reco_utils to support common tasks such as loading datasets in the format expected by different algorithms, evaluating model outputs, and splitting train/test data. Implementations of several state-of-the-art algorithms are provided for self-study and customization in your own applications.

Getting Started

Please see the setup guide for more details on setting up your machine locally, on Spark, or on Azure Databricks.

To setup on your local machine:

  1. Install Anaconda with Python >= 3.6. Miniconda is a quick way to get started.
  2. Clone the repository
    git clone
  3. Run the generate conda file script and create a conda environment:
    cd Recommenders
    conda env create -n reco -f conda_bare.yaml  
  4. Activate the conda environment and register it with Jupyter:
    conda activate reco
    python -m ipykernel install --user --name reco --display-name "Python (reco)"
  5. Start the Jupyter notebook server
    cd notebooks
    jupyter notebook
  6. Run the SAR Python CPU Movielens notebook under the 00_quick_start folder. Make sure to change the kernel to "Python (reco)".


We provide several notebooks to show how recommendation algorithms can be designed, evaluated and operationalized.

The Quick-Start and Modeling notebooks showcase how to utilize the following algorithms to build a recommender system:


Algorithm Environment Type Description
Classic Recommenders
Surprise/Singular Value Decomposition (SVD) Python Collaborative Filtering General purpose algorithm for smaller datasets
Alternating Least Squares (ALS) Spark Collaborative General purpose algorithm for larger datasets, optimized with Spark
Microsoft Recommenders
Smart Adaptive Recommendations (SAR) Python / Spark Collaborative Filtering Generalized algorithm utilizing item similarities and can easily adapt to new users
Vowpal Wabbit Family (VW) Python / Online Collaborative, Content Based Fast online learning algorithms, great for scenarios where user features / context are constantly changing, like real-time bidding
eXtreme Deep Factorization Machine (xDeepFM) Python / GPU Hybrid Deep learning model combining implicit and explicit features
Deep Knowledge-Aware Network (DKN) Python / GPU Content Based Deep learning model incorporating a knowledge graph and article embeddings to provide powerful news or article recommendations
Deep Learning
Neural Collaborative Filtering (NCF) Python / GPU Collaborative Filtering General algorithm built using a multi-layer perceptron
Restricted Boltzmann Machines (RBM) Python / GPU Collaborative Filtering Generative neural network algorithm built to learn the underlying probability distribution for user/item affinity
FastAI Embedding Dot Bias (FAST) Python / GPU Collaborative Filtering General purpose algorithm embedding dot biases for users and items

In addition, we also provide a comparison notebook to illustrate how different algorithms could be evaluated and compared. In this notebook, data (MovieLens 1M) is randomly split into train/test sets at a 75/25 ratio. A recommendation model is trained using each of the collaborative filtering algorithms below. We utilize empirical parameter values reported in literature here. For ranking metrics we use k = 10 (top 10 results). We run the comparison on a Standard NC6s_v2 Azure DSVM (6 vCPUs, 112 GB memory and 1 K80 GPU). Spark ALS is run in local standalone mode.

Preliminary Comparison

Algo MAP nDCG@k Precision@k Recall@k RMSE MAE R2 Explained Variance
ALS 0.002020 0.024313 0.030677 0.009649 0.860502 0.680608 0.406014 0.411603
SVD 0.010915 0.102398 0.092996 0.025362 0.888991 0.696781 0.364178 0.364178
FastAI 0.023022 0.168714 0.154761 0.050153 0.887224 0.705609 0.371552 0.374281


This project welcomes contributions and suggestions. Before contributing, please see our contribution guidelines.

Build Status

Build Type Branch Status Branch Status
Linux CPU master Status staging Status
Linux GPU master Status staging Status
Linux Spark master Status staging Status

NOTE - the tests are executed every night, we use pytest for testing python utilities in reco_utils and papermill for the notebooks.