Fast and memory efficient library for factorization machines (FM).

• Supports both ℓ1 regularized logistic regression and factorization machines.
• Runs on local machine and distributed clusters.
• Scales to datasets with billions examples and features.

The following commands clone and build difacto, then download a sample dataset, and train FM with 2-dimension on it.

git clone --recursive https://github.com/dmlc/difacto
cd difacto; git submodule update --init; make -j8
./tools/download.sh gisette
build/difacto data_in=data/gisette_scale val_data=data/gisette_scale.t lr=.02 V_dim=2 V_lr=.001

Origins from wormhole/learn/difacto.

(NOTE: this project is still under developing)

Mu Li, Ziqi Liu, Alex Smola, and Yu-Xiang Wang. DiFacto — Distributed Factorization Machines. In WSDM, 2016