Build Status

Fast Python Collaborative Filtering for Implicit Datasets.

This project provides fast Python implementations of several different popular recommendation algorithms for implicit feedback datasets:

All models have multi-threaded training routines, using Cython and OpenMP to fit the models in parallel among all available CPU cores. In addition, the ALS and BPR models both have custom CUDA kernels – enabling fitting on compatible GPU’s. Approximate nearest neighbours libraries such as Annoy, NMSLIB and Faiss can also be used by Implicit to speed up making recommendations.


There are binary packages on conda-forge for Linux, Windows and OSX. These can be installed with:

conda install -c conda-forge implicit

There are also GPU enabled packages on conda-forge for x86_64 Linux systems using either CUDA 9.2, 10.0, 10.1 or 10.2. The GPU packages can be installed with:

conda install -c conda-forge implicit implicit-proc=*=gpu

There is also an sdist package on PyPi. This package can be installed with:

pip install implicit

Note that installing with pip requires a C++ compiler to be installed on your system, since this method will build implicit from source.

Basic Usage

import implicit

# initialize a model
model = implicit.als.AlternatingLeastSquares(factors=50)

# train the model on a sparse matrix of item/user/confidence weights

# recommend items for a user
user_items = item_user_data.T.tocsr()
recommendations = model.recommend(userid, user_items)

# find related items
related = model.similar_items(itemid)

The examples folder has a program showing how to use this to compute similar artists on the dataset.

For more information see the documentation.

Articles about Implicit

These blog posts describe the algorithms that power this library:

There are also several other blog posts about using Implicit to build recommendation systems:


This library requires SciPy version 0.16 or later. Running on OSX requires an OpenMP compiler, which can be installed with homebrew: brew install gcc. Running on Windows requires Python 3.5+.

GPU Support requires at least version 9 of the NVidia CUDA Toolkit. The build will use the nvcc compiler that is found on the path, but this can be overriden by setting the CUDAHOME enviroment variable to point to your cuda installation.

This library has been tested with Python 2.7, 3.5, 3.6 and 3.7 on Ubuntu and OSX, and tested with Python 3.5 and 3.6 on Windows.


Simple benchmarks comparing the ALS fitting time versus Spark and QMF can be found here.

Optimal Configuration

I’d recommend configuring SciPy to use Intel’s MKL matrix libraries. One easy way of doing this is by installing the Anaconda Python distribution.

For systems using OpenBLAS, I highly recommend setting ‘export OPENBLAS_NUM_THREADS=1’. This disables its internal multithreading ability, which leads to substantial speedups for this package. Likewise for Intel MKL, setting ‘export MKL_NUM_THREADS=1’ should also be set.

Released under the MIT License