libffm

A Python wrapper for the libffm library.

Quick start

git clone git@github.com:turi-code/GraphLab-Create-SDK.git sdk
git clone git@github.com:turi-code/python-libffm.git ffm
cd ffm
make

To run the following examples you will also need to register for GraphLab Create. This software is free for non-commercial use and has a 30 day free trial otherwise.

After that, try running the basic example:

ipython examples/basic.py

If you want to try a less synthetic example, download the 1TB Criteo dataset. First test things out with a small sample of the dataset.

gzip -cd day_0.gz| head -n 1000000 > criteo-sample.tsv

Next we have a sample script for performing some of the same types of feature engineering that the contest winners have been using:

ipython examples/criteo_process.py

Train a FFM model on this data.

ipython examples/criteo_sample.py

You should see something like the following (which appears to be overfitting in this example):

PROGRESS: iter   tr_logloss   va_logloss
PROGRESS:    0      0.12794      0.12353
PROGRESS:    1      0.10907      0.12636
PROGRESS:    2      0.09263      0.13318
PROGRESS:    3      0.07679      0.14200
PROGRESS:    4      0.06411      0.15130
PROGRESS:    5      0.05484      0.16034
...

Usage

The package makes it easy to train models directly from SFrames.

import ffm

train = gl.SFrame('examples/small.tr.sframe')
test = gl.SFrame('examples/small.te.sframe')

m = ffm.FFM(lam=.1)
m.fit(train, target='y', nr_iters=50)
yhat = m.predict(test)

Each column is interpreted as a separate "field" in the model. Only dict columns are currently supported, where the keys of each dict are integers that represent the feature id.

Code

• libfmm.cpp: uses C++ macros provided by Turi's SDK to wrap libffm's methods as Python classes and methods.
• fmm.py: a scikit-learn-style wrapper.
• lib/: the original library, where cout statements have been replaced with Turi's progress_stream to allow progress printing to Python.
• examples/: example scripts for training models using the sample data provided with the original package as well as with data similar to Kaggle's criteo competition.

More details

For more on how and why we made this, see the blog post.

License

This package provided under the 3-clause BSD license.