README

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:

<label> <field1>:<feature1>:<value1> <field2>:<feature2>:<value2> ...
.
.
.

`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 <lambda>: set regularization parameter (default 0.00002)
    -k <factor>: set number of latent factors (default 4)
    -t <iteration>: set number of iterations (default 15)
    -r <eta>: set learning rate (default 0.2)
    -s <nr_threads>: set number of threads (default 1)
    -p <path>: 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://www.csie.ntu.edu.tw/~f01922139/libffm_data/libffm_toy.zip

    tar.gz: https://www.csie.ntu.edu.tw/~f01922139/libffm_data/libffm_toy.tar.gz

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

For questions, comments, feature requests, or bug report, please send your email to:

    Yuchin Juan (juan.yuchin@gmail.com)

For Windows related questions, please send your email to:

    Wei-Sheng Chin (image.chin@gmail.com)