Sparse LInear Method (SLIM) [1] is an item-based top-N recommendation approach that combines the advantages of neighborhood- and model-based collaborative filtering methods. It achieves state-of-the-art recommendation performance and has low computational requirements.
This package provides a C-based optimized multi-threaded implementation of SLIM that consists of a set of command-line programs and a user-callable library for estimating and applying SLIM models as well as an easy to use Python interface.
SLIM uses Git submodules to manage external dependencies. Hence, please specify the --recursive option while cloning the repo as follow:
git clone --recursive https://github.com/KarypisLab/SLIM.gitTo build SLIM you can follow the instructions below:
General dependencies for building slim are: gcc, cmake, build-essential. In Ubuntu systems these can be obtained from the apt package manager (e.g., apt-get install cmake, etc)
sudo apt-get install build-essential
sudo apt-get install cmakeIn order to build SLIM, first build GKlib by running:
cd lib/GKlib
make config openmp=set
make
cd ../../After building GKlib, you can build and install SLIM by running:
make config shared=1 cc=gcc cxx=gcc prefix=~/.local
make installIn order to use SLIM's ADMM solver, you will need to install Intel's MKL library.
For Ubuntu machines on which you have sudo privileges, we provided the depmkl.sh script that automates the process of obtaining and installing MKL, which can be used as follows:
bash depmkl.sh
source ~/.bashrc For machines on which you do not have sudo privileges, you should download the MKL tarball from Intel's website and then install it locally using the install.sh script they provide. After installing it you should add your-path-to-intel/intel/mkl/bin/mklvars.sh intel64in your bashrc and run source ~/.bashrc.
Next you can build and install SLIM with MKL support by running:
make config shared=1 cc=gcc cxx=gcc with_mkl=1 prefix=~/.local
make installNote that SLIM's ADMM solver usually outperforms the default optimizer included in SLIM when the number of items in the dataset is relatively small compared to the number of users and the number of non-zeros in the dataset is large.
The Python package is located at python-package/.
The installation of python-package requires Python distutils module and is often part of the core Python package or can be installed using a package manager, e.g., in Debian use
sudo apt-get install python-setuptoolsAfter building the SLIM library, follow one of the following steps to install the python-package:
- Install the python-package system-wide (this requires sudo privileges):
cd python-package
sudo python setup.py install- Install the python-package only for the current user (without sudo privileges):
cd python-package
python setup.py install --userHere are some examples to quickly try out SLIM on the sample datasets that are provided with SLIM.
import pandas as pd
from SLIM import SLIM, SLIMatrix
#read training data stored as triplets <user> <item> <rating>
traindata = pd.read_csv('../test/AutomotiveTrain.ijv', delimiter = ' ', header=None)
trainmat = SLIMatrix(traindata)
#set up parameters to learn model, e.g., use Coordinate Descent with L1 and L2
#regularization
params = {'algo':'cd', 'nthreads':2, 'l1r':1.0, 'l2r':1.0}
#learn the model using training data and desired parameters
model = SLIM()
model.train(params, trainmat)
#read test data having candidate items for users
testdata = pd.read_csv('../test/AutomotiveTest.ijv', delimiter = ' ', header=None)
#NOTE: model object is passed as an argument while generating test matrix
testmat = SLIMatrix(testdata, model)
#generate top-10 recommendations
prediction_res = model.predict(testmat, nrcmds=10, outfile = 'output.txt')
#dump the model to files on disk
model.save_model(modelfname='model.csr', # filename to save the model as a csr matrix
mapfname='map.csr' # filename to save the item map
)
#load the model from from disk
model_new = SLIM()
model_new.load_model(modelfname='model.csr', # filename of the model
mapfname='map.csr' # filename of the item map
)The users can also refer to the python notebook UserGuide.ipynb located at
./python-package/UserGuide.ipynb for more examples on using the python api.
SLIM can be used by running the command-line programs that are located under ./build directory. Specifically, SLIM provides the following three command-line programs:
slim_learn: for estimating a modelslim_predict: for applying a previously estimated model, andslim_mselect: for exploring a set of hyper-parameters in order to select the best performing model.
Additional information about how to use these command-line programs is located in SLIM's reference manual that is available at ./doxygen/html/index.html or ./doxygen/latex/refman.pdf.
You can also use SLIM by direclty linking into your C/C++ program via its library interface. SLIM's API is described in SLIM's reference manual (see links above).
If you use any part of this library in your research, please cite it using the following BibTex entry:
@online{slim,
title = {{SLIM Library for Recommender Systems}},
author = {Ning, Xia and Nikolakopoulos, Athanasios N. and Shui, Zeren and Sharma, Mohit and Karypis, George},
url = {https://github.com/KarypisLab/SLIM},
year = {2019},
}
This implementation of SLIM was written by George Karypis with contributions by Xia Ning, Athanasios N. Nikolakopoulos, Zeren Shui and Mohit Sharma.
If you encounter any problems or have any suggestions, please contact George Karypis at [email protected].
Copyright 2019, Regents of the University of Minnesota
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.