This repository contains 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 classical and deep learning recommender algorithms such as Alternating Least Squares (ALS) or eXtreme Deep Factorization Machines (xDeepFM).
- Evaluate: Evaluating algorithms with offline metrics
- Model Select and Optimize: Tuning and optimizing hyperparameters 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 training/test data. Implementations of several state-of-the-art algorithms are included for self-study and customization in your own applications. See the reco_utils documentation.
For a more detailed overview of the repository, please see the documents at the wiki page.
Please see the setup guide for more details on setting up your machine locally, on a data science virtual machine (DSVM) or on Azure Databricks.
To setup on your local machine:
-
Install Anaconda with Python >= 3.6. Miniconda is a quick way to get started.
-
Clone the repository
git clone https://github.com/Microsoft/Recommenders
- Run the generate conda file script to create a conda environment: (This is for a basic python environment, see SETUP.md for PySpark and GPU environment setup)
cd Recommenders
python scripts/generate_conda_file.py
conda env create -f reco_base.yaml
- Activate the conda environment and register it with Jupyter:
conda activate reco_base
python -m ipykernel install --user --name reco_base --display-name "Python (reco)"
- Start the Jupyter notebook server
jupyter notebook
- Run the SAR Python CPU MovieLens notebook under the
00_quick_startfolder. Make sure to change the kernel to "Python (reco)".
NOTE - The Alternating Least Squares (ALS) notebooks require a PySpark environment to run. Please follow the steps in the setup guide to run these notebooks in a PySpark environment. For the deep learning algorithms, it is recommended to use a GPU machine.
The table below lists the recommender algorithms currently available in the repository. Notebooks are linked under the Environment column when different implementations are available.
| Algorithm | Environment | Type | Description |
|---|---|---|---|
| Alternating Least Squares (ALS) | PySpark | Collaborative Filtering | Matrix factorization algorithm for explicit or implicit feedback in large datasets, optimized by Spark MLLib for scalability and distributed computing capability |
| Attentive Asynchronous Singular Value Decomposition (A2SVD)* | Python CPU / Python GPU | Collaborative Filtering | Sequential-based algorithm that aims to capture both long and short-term user preferences using attention mechanism |
| Cornac/Bayesian Personalized Ranking (BPR) | Python CPU | Collaborative Filtering | Matrix factorization algorithm for predicting item ranking with implicit feedback |
| Convolutional Sequence Embedding Recommendation (Caser) | Python CPU / Python GPU | Collaborative Filtering | Algorithm based on convolutions that aims to capture both user’s general preferences and sequential patterns |
| Deep Knowledge-Aware Network (DKN)* | Python CPU / Python GPU | Content-Based Filtering | Deep learning algorithm incorporating a knowledge graph and article embeddings to provide powerful news or article recommendations |
| Extreme Deep Factorization Machine (xDeepFM)* | Python CPU / Python GPU | Hybrid | Deep learning based algorithm for implicit and explicit feedback with user/item features |
| FastAI Embedding Dot Bias (FAST) | Python CPU / Python GPU | Collaborative Filtering | General purpose algorithm with embeddings and biases for users and items |
| LightGBM/Gradient Boosting Tree* | Python CPU / PySpark | Content-Based Filtering | Gradient Boosting Tree algorithm for fast training and low memory usage in content-based problems |
| GRU4Rec | Python CPU / Python GPU | Collaborative Filtering | Sequential-based algorithm that aims to capture both long and short-term user preferences using recurrent neural networks |
| Neural Recommendation with Long- and Short-term User Representations (LSTUR)* | Python CPU / Python GPU | Content-Based Filtering | Neural recommendation algorithm with long- and short-term user interest modeling |
| Neural Recommendation with Attentive Multi-View Learning (NAML)* | Python CPU / Python GPU | Content-Based Filtering | Neural recommendation algorithm with attentive multi-view learning |
| Neural Collaborative Filtering (NCF) | Python CPU / Python GPU | Collaborative Filtering | Deep learning algorithm with enhanced performance for implicit feedback |
| Neural Recommendation with Personalized Attention (NPA)* | Python CPU / Python GPU | Content-Based Filtering | Neural recommendation algorithm with personalized attention network |
| Neural Recommendation with Multi-Head Self-Attention (NRMS)* | Python CPU / Python GPU | Content-Based Filtering | Neural recommendation algorithm with multi-head self-attention |
| Restricted Boltzmann Machines (RBM) | Python CPU / Python GPU | Collaborative Filtering | Neural network based algorithm for learning the underlying probability distribution for explicit or implicit feedback |
| Riemannian Low-rank Matrix Completion (RLRMC)* | Python CPU | Collaborative Filtering | Matrix factorization algorithm using Riemannian conjugate gradients optimization with small memory consumption. |
| Simple Algorithm for Recommendation (SAR)* | Python CPU | Collaborative Filtering | Similarity-based algorithm for implicit feedback dataset |
| Short-term and Long-term preference Integrated Recommender (SLi-Rec)* | Python CPU / Python GPU | Collaborative Filtering | Sequential-based algorithm that aims to capture both long and short-term user preferences using attention mechanism, a time-aware controller and a content-aware controller |
| Surprise/Singular Value Decomposition (SVD) | Python CPU | Collaborative Filtering | Matrix factorization algorithm for predicting explicit rating feedback in datasets that are not very large |
| Term Frequency - Inverse Document Frequency (TF-IDF) | Python CPU | Content-Based Filtering | Simple similarity-based algorithm for content-based recommendations with text datasets |
| Vowpal Wabbit Family (VW)* | Python CPU (online training) | Content-Based Filtering | Fast online learning algorithms, great for scenarios where user features / context are constantly changing |
| Wide and Deep | Python CPU / Python GPU | Hybrid | Deep learning algorithm that can memorize feature interactions and generalize user features |
| xLearn/Factorization Machine (FM) & Field-Aware FM (FFM) | Python CPU | Content-Based Filtering | Quick and memory efficient algorithm to predict labels with user/item features |
NOTE: * indicates algorithms invented/contributed by Microsoft.
Independent or incubating algorithms and utilities are candidates for the contrib folder. This will house contributions which may not easily fit into the core repository or need time to refactor or mature the code and add necessary tests.
| Algorithm | Environment | Type | Description |
|---|---|---|---|
| SARplus * | PySpark | Collaborative Filtering | Optimized implementation of SAR for Spark |
We provide a benchmark notebook to illustrate how different algorithms could be evaluated and compared. In this notebook, the MovieLens dataset is split into training/test sets at a 75/25 ratio using a stratified split. 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 recommended items). We run the comparison on a Standard NC6s_v2 Azure DSVM (6 vCPUs, 112 GB memory and 1 P100 GPU). Spark ALS is run in local standalone mode. In this table we show the results on Movielens 100k, running the algorithms for 15 epochs.
| Algo | MAP | nDCG@k | Precision@k | Recall@k | RMSE | MAE | R2 | Explained Variance |
|---|---|---|---|---|---|---|---|---|
| ALS | 0.004732 | 0.044239 | 0.048462 | 0.017796 | 0.965038 | 0.753001 | 0.255647 | 0.251648 |
| SVD | 0.012873 | 0.095930 | 0.091198 | 0.032783 | 0.938681 | 0.742690 | 0.291967 | 0.291971 |
| SAR | 0.113028 | 0.388321 | 0.333828 | 0.183179 | N/A | N/A | N/A | N/A |
| NCF | 0.107720 | 0.396118 | 0.347296 | 0.180775 | N/A | N/A | N/A | N/A |
| BPR | 0.105365 | 0.389948 | 0.349841 | 0.181807 | N/A | N/A | N/A | N/A |
| FastAI | 0.025503 | 0.147866 | 0.130329 | 0.053824 | 0.943084 | 0.744337 | 0.285308 | 0.287671 |
This project welcomes contributions and suggestions. Before contributing, please see our contribution guidelines.
These tests are the nightly builds, which compute the smoke and integration tests. master is our main branch and staging is our development branch. We use pytest for testing python utilities in reco_utils and papermill for the notebooks. For more information about the testing pipelines, please see the test documentation.
The following tests run on a Windows and Linux DSVM daily. These machines run 24/7.
| Build Type | Branch | Status | Branch | Status | |
|---|---|---|---|---|---|
| Linux CPU | master | staging | |||
| Linux GPU | master | staging | |||
| Linux Spark | master | staging | |||
| Windows CPU | master | staging | |||
| Windows GPU | master | staging | |||
| Windows Spark | master | staging |
Microsoft AI Github: Find other Best Practice projects, and Azure AI design patterns in our central repository.