Table of Contents

• What is LIBFFM
• Overfitting and Early Stopping
• Installation
• Data Format
• Command Line Usage
• Examples
• OpenMP and SSE
• Building Windows Binaries
• FAQ

What is LIBFFM

LIBFFM is a library for field-aware factorization machine (FFM).

Field-aware factorization machine is a effective model for CTR prediction. It has been used to win the top-3 positions of following competitions:

* Criteo: https://www.kaggle.com/c/criteo-display-ad-challenge

* Avazu: https://www.kaggle.com/c/avazu-ctr-prediction

* Outbrain: https://www.kaggle.com/c/outbrain-click-prediction

* RecSys 2015: http://dl.acm.org/citation.cfm?id=2813511&dl=ACM&coll=DL&CFID=941880276&CFTOKEN=60022934

You can find more information about FFM in the following paper / slides:

* http://www.csie.ntu.edu.tw/~r01922136/slides/ffm.pdf

* http://www.csie.ntu.edu.tw/~cjlin/papers/ffm.pdf

* https://arxiv.org/abs/1701.04099

Overfitting and Early Stopping

FFM is prone to overfitting, and the solution we have so far is early stopping. See how FFM behaves on a certain data set:

> ffm-train -p va.ffm -l 0.00002 tr.ffm
iter   tr_logloss   va_logloss
   1      0.49738      0.48776
   2      0.47383      0.47995
   3      0.46366      0.47480
   4      0.45561      0.47231
   5      0.44810      0.47034
   6      0.44037      0.47003
   7      0.43239      0.46952
   8      0.42362      0.46999
   9      0.41394      0.47088
  10      0.40326      0.47228
  11      0.39156      0.47435
  12      0.37886      0.47683
  13      0.36522      0.47975
  14      0.35079      0.48321
  15      0.33578      0.48703

We see the best validation loss is achieved at 7th iteration. If we keep training, then overfitting begins. It is worth noting that increasing regularization parameter do not help:

> ffm-train -p va.ffm -l 0.0002 -t 50 -s 12 tr.ffm
iter   tr_logloss   va_logloss
   1      0.50532      0.49905
   2      0.48782      0.49242
   3      0.48136      0.48748
             ...
  29      0.42183      0.47014
             ...
  48      0.37071      0.47333
  49      0.36767      0.47374
  50      0.36472      0.47404

To avoid overfitting, we recommend always provide a validation set with option -p.' You can use option --auto-stop' to stop at the iteration that reaches the best validation loss:

> ffm-train -p va.ffm -l 0.00002 --auto-stop tr.ffm
iter   tr_logloss   va_logloss
   1      0.49738      0.48776
   2      0.47383      0.47995
   3      0.46366      0.47480
   4      0.45561      0.47231
   5      0.44810      0.47034
   6      0.44037      0.47003
   7      0.43239      0.46952
   8      0.42362      0.46999
Auto-stop. Use model at 7th iteration.

Installation

Requirement: It requires a C++11 compatible compiler. We also use OpenMP to provide multi-threading. If OpenMP is not available on your platform, please refer to section `OpenMP and SSE.'

• Unix-like systems: Typeype `make' in the command line.

• Windows: See `Building Windows Binaries' to compile.

Data Format

The data format of LIBFFM is:

:: :: ... . . .

field' and feature' should be non-negative integers. See an example `bigdata.tr.txt.'

It is important to understand the difference between field' and feature'. For example, if we have a raw data like this:

Click Advertiser Publisher ===== ========== ========= 0 Nike CNN 1 ESPN BBC

Here, we have

* 2 fields: Advertiser and Publisher

* 4 features: Advertiser-Nike, Advertiser-ESPN, Publisher-CNN, Publisher-BBC

Usually you will need to build two dictionares, one for field and one for features, like this:

DictField[Advertiser] -> 0
DictField[Publisher]  -> 1

DictFeature[Advertiser-Nike] -> 0
DictFeature[Publisher-CNN]   -> 1
DictFeature[Advertiser-ESPN] -> 2
DictFeature[Publisher-BBC]   -> 3

Then, you can generate FFM format data:

0 0:0:1 1:1:1
1 0:2:1 1:3:1

Note that because these features are categorical, the values here are all ones.

Command Line Usage

• `ffm-train'

  usage: ffm-train [options] training_set_file [model_file]

  options: -l : set regularization parameter (default 0.00002) -k : set number of latent factors (default 4) -t : set number of iterations (default 15) -r : set learning rate (default 0.2) -s <nr_threads>: set number of threads (default 1) -p : set path to the validation set --quiet: quiet model (no output) --no-norm: disable instance-wise normalization --auto-stop: stop at the iteration that achieves the best validation loss (must be used with -p)

  By default we do instance-wise normalization. That is, we normalize the 2-norm of each instance to 1. You can use `--no-norm' to disable this function.

  A binary file `training_set_file.bin' will be generated to store the data in binary format.

  Because FFM usually need early stopping for better test performance, we provide an option --auto-stop' to stop at the iteration that achieves the best validation loss. Note that you need to provide a validation set with -p' when you use this option.

• `ffm-predict'

  usage: ffm-predict test_file model_file output_file

Examples

Download a toy data from:

zip: https://drive.google.com/open?id=1HZX7zSQJy26hY4_PxSlOWz4x7O-tbQjt

tar.gz: https://drive.google.com/open?id=12-EczjiYGyJRQLH5ARy1MXRFbCvkgfPx

This dataset is subsampled 1% from Criteo's challenge.

tar -xzf libffm_toy.tar.gz

or

unzip libffm_toy.zip

./ffm-train -p libffm_toy/criteo.va.r100.gbdt0.ffm libffm_toy/criteo.tr.r100.gbdt0.ffm model

train a model using the default parameters

./ffm-predict libffm_toy/criteo.va.r100.gbdt0.ffm model output

do prediction

./ffm-train -l 0.0001 -k 15 -t 30 -r 0.05 -s 4 --auto-stop -p libffm_toy/criteo.va.r100.gbdt0.ffm libffm_toy/criteo.tr.r100.gbdt0.ffm model

train a model using the following parameters:

regularization cost = 0.0001
latent factors = 15
iterations = 30
learning rate = 0.3
threads = 4
let it auto-stop

OpenMP and SSE

We use OpenMP to do parallelization. If OpenMP is not available on your platform, then please comment out the following lines in Makefile.

DFLAG += -DUSEOMP
CXXFLAGS += -fopenmp

Note: Please run `make clean all' if these flags are changed.

We use SSE instructions to perform fast computation. If you do not want to use it, comment out the following line:

DFLAG += -DUSESSE

Then, run `make clean all'

Building Windows Binaries

The Windows part is maintained by different maintainer, so it may not always support the latest version.

The latest version it supports is: v1.21

To build them via command-line tools of Visual C++, use the following steps:

1. Open a DOS command box (or Developer Command Prompt for Visual Studio) and go to LIBFFM directory. If environment variables of VC++ have not been set, type

"C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\bin\amd64\vcvars64.bat"

You may have to modify the above command according which version of VC++ or where it is installed.

2. Type

nmake -f Makefile.win clean all

FAQ

Q: Why I have the same model size when k = 1 and k = 4?

A: This is because we use SSE instructions. In order to use SSE, the memory need to be aligned. So even you assign k = 1, we still fill some dummy zeros from k = 2 to 4.

Q: Why the logloss is slightly different on the same data when I run the program two or more times when I use multi-threading

A: When there are more then one thread, the program becomes non-deterministic. To make it determinisitc you can only use one thread.

Contributors

Yuchin Juan, Wei-Sheng Chin, and Yong Zhuang