pywFM
pywFM is a Python wrapper for Steffen Rendle's libFM. libFM is a Factorization Machine library:
Factorization machines (FM) are a generic approach that allows to mimic most factorization models by feature engineering. This way, factorization machines combine the generality of feature engineering with the superiority of factorization models in estimating interactions between categorical variables of large domain. libFM is a software implementation for factorization machines that features stochastic gradient descent (SGD) and alternating least squares (ALS) optimization as well as Bayesian inference using Markov Chain Monte Carlo (MCMC).
For more information regarding Factorization machines and libFM, read Steffen Rendle's paper: Factorization Machines with libFM, in ACM Trans. Intell. Syst. Technol., 3(3), May. 2012
Don't forget to acknowledge libFM (i.e. cite the paper Factorization Machines with libFM) if you publish results produced with this software.
Motivation
While using Python implementations of Factorization Machines, I felt that the current implementations (pyFM and fastFM) had many flaws. Then I though, why re-invent the wheel? Why not use the original libFM?
Sure, it's not Python native yada yada ... But at least we have a bulletproof, battle-tested implementation that we can guide ourselves with.
Installing
First you have to clone and compile libFM repository and set an environment variable to the libFM bin folder:
git clone https://github.com/srendle/libfm /home/libfm
cd /home/libfm/
# taking advantage of a bug to allow us to save model #ShameShame
git reset --hard 91f8504a15120ef6815d6e10cc7dee42eebaab0f
make all
export LIBFM_PATH=/home/libfm/bin/
Make sure you are compiling source from libfm repository and at this specific commit, since pywFM needs the save_model. Beware that the installers and source code in libfm.org are both dated before this commit. I know this is extremely hacky, but since a fix was deployed it only allows the save_model option for SGD or ALS. I don't know why exactly, because it was working well before.
If you use Jupyter take a look at the following issue for some extra notes on getting libfm to work.
Then, install pywFM using pip:
pip install pywFM
Binary installers for the latest released version are available at the Python package index.
Dependencies
- numpy
- scipy
- sklearn
- pandas
Example
Very simple example taken from Steffen Rendle's paper: Factorization Machines with libFM.
import pywFM
import numpy as np
import pandas as pd
features = np.matrix([
# Users | Movies | Movie Ratings | Time | Last Movies Rated
# A B C | TI NH SW ST | TI NH SW ST | | TI NH SW ST
[1, 0, 0, 1, 0, 0, 0, 0.3, 0.3, 0.3, 0, 13, 0, 0, 0, 0 ],
[1, 0, 0, 0, 1, 0, 0, 0.3, 0.3, 0.3, 0, 14, 1, 0, 0, 0 ],
[1, 0, 0, 0, 0, 1, 0, 0.3, 0.3, 0.3, 0, 16, 0, 1, 0, 0 ],
[0, 1, 0, 0, 0, 1, 0, 0, 0, 0.5, 0.5, 5, 0, 0, 0, 0 ],
[0, 1, 0, 0, 0, 0, 1, 0, 0, 0.5, 0.5, 8, 0, 0, 1, 0 ],
[0, 0, 1, 1, 0, 0, 0, 0.5, 0, 0.5, 0, 9, 0, 0, 0, 0 ],
[0, 0, 1, 0, 0, 1, 0, 0.5, 0, 0.5, 0, 12, 1, 0, 0, 0 ]
])
target = [5, 3, 1, 4, 5, 1, 5]
fm = pywFM.FM(task='regression', num_iter=5)
# split features and target for train/test
# first 5 are train, last 2 are test
model = fm.run(features[:5], target[:5], features[5:], target[5:])
print(model.predictions)
# you can also get the model weights
print(model.weights)
You can also use numpy's array, sklearn's sparse_matrix, and even pandas' DataFrame as features input.
Prediction on new data
Current approach is to send test data as x_test, y_test in run method call. libfm uses the test values to output some results regarding its predictions. They are not used when training the model. y_test can be set as dummy value and just collect the predictions with model.predictions (also disregard the prediction statistics since those will be wrong). For more info check libfm manual.
Running against a new dataset using something like a predict method is not supported yet. Pending feature request: https://github.com/jfloff/pywFM/issues/7
Feel free to PR that change ;)
Usage
Don't forget to acknowledge libFM (i.e. cite the paper Factorization Machines with libFM) if you publish results produced with this software.
FM: Class that wraps libFM parameters. For more information read libFM manual
Parameters
----------
task : string, MANDATORY
regression: for regression
classification: for binary classification
num_iter: int, optional
Number of iterations
Defaults to 100
init_stdev : double, optional
Standard deviation for initialization of 2-way factors
Defaults to 0.1
k0 : bool, optional
Use bias.
Defaults to True
k1 : bool, optional
Use 1-way interactions.
Defaults to True
k2 : int, optional
Dimensionality of 2-way interactions.
Defaults to 8
learning_method: string, optional
sgd: parameter learning with SGD
sgda: parameter learning with adpative SGD
als: parameter learning with ALS
mcmc: parameter learning with MCMC
Defaults to 'mcmc'
learn_rate: double, optional
Learning rate for SGD
Defaults to 0.1
r0_regularization: int, optional
bias regularization for SGD and ALS
Defaults to 0
r1_regularization: int, optional
1-way regularization for SGD and ALS
Defaults to 0
r2_regularization: int, optional
2-way regularization for SGD and ALS
Defaults to 0
rlog: bool, optional
Enable/disable rlog output
Defaults to True.
verbose: bool, optional
How much infos to print
Defaults to False.
seed: int, optional
seed used to reproduce the results
Defaults to None.
silent: bool, optional
Completly silences all libFM output
Defaults to False.
temp_path: string, optional
Sets path for libFM temporary files. Usefull when dealing with large data.
Defaults to None (default mkstemp behaviour)
FM.run: run factorization machine model against train and test data
Parameters
----------
x_train : {array-like, matrix}, shape = [n_train, n_features]
Training data
y_train : numpy array of shape [n_train]
Target values
x_test: {array-like, matrix}, shape = [n_test, n_features]
Testing data
y_test : numpy array of shape [n_test]
Testing target values
x_validation_set: optional, {array-like, matrix}, shape = [n_train, n_features]
Validation data (only for SGDA)
y_validation_set: optional, numpy array of shape [n_train]
Validation target data (only for SGDA)
Return
-------
Returns `namedtuple` with the following properties:
predictions: array [n_samples of x_test]
Predicted target values per element in x_test.
global_bias: float
If k0 is True, returns the model's global bias w0
weights: array [n_features]
If k1 is True, returns the model's weights for each features Wj
pairwise_interactions: numpy matrix [n_features x k2]
Matrix with pairwise interactions Vj,f
rlog: pandas dataframe [nrow = num_iter]
`pandas` DataFrame with measurements about each iteration
Docker
This repository includes Dockerfile for development and for running pywFM.
- Run
pywFMexamples (Dockerfile): if you are only interested in running the examples, you can use the pre-build image availabe in Docker Hub:
# to run examples/simple.py (the one in this README).
docker run --rm -v "$(pwd)":/home/pywfm -w /home/pywfm -ti jfloff/pywfm python examples/simple.py
- Development of
pywFM(Dockerfile): useful if you want to make changes to the repo.Dockerfiledefaults to bash.
# to build image
docker build --rm=true -t jfloff/pywfm-dev .
# to run image
docker run --rm -v "$(pwd)":/home/pywfm-dev -w /home/pywfm-dev -ti jfloff/pywfm-dev
Future work
- Improve the
save_model/load_modelso we can have a more defined init-fit-predict cycle (perhaps we could inherit from sklearn.BaseEstimator) - Can we contribute to libFM repo so save_model is enabled for all learning methods (namely MCMC)?
- Look up into shared library solution to improve I/O overhead
I'm no factorization machine expert, so this library was just an effort to have libFM as fast as possible in Python. Feel free to suggest features, enhancements; to point out issues; and of course, to post PRs.
License
MIT (see LICENSE.txt file)
892 Jan 3, 2023
785 Dec 21, 2022
174 Dec 7, 2022
757 Dec 31, 2022
23.3k Dec 31, 2022
6 Jun 30, 2022
4.2k Dec 29, 2022
2k Jan 2, 2023
1.4k Jan 6, 2023
23.6k Jan 3, 2023
6.9k Jan 5, 2023
1.3k Dec 22, 2022
5.2k Jan 4, 2023
2.5k Jan 6, 2023
696 Dec 26, 2022
438 Dec 17, 2022
1.6k Dec 31, 2022
152 Jan 2, 2023
14 Dec 19, 2021