LIBFFM is a library for field-aware factorization machine. For the formulation it solves, please check:
http://www.csie.ntu.edu.tw/~r01922136/slides/ffm.pdf
Table of Contents
=================
- Overfitting and Early Stopping
- Installation
- Data Format
- Command Line Usage
- Examples
- Library Usage
- OpenMP
- Building Windows Binaries
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: LIBFFM is written in C++. It requires C++11 and OpenMP supports. If OpenMP is not available on your
platform, please refer to section `OpenMP.'
- Unix-like systems:
To compile on Unix-like systems, type `make' in the command line.
- OS X:
The built-in compiler should be able to compile LIBFFM. However, OpenMP may
not be supported. In this case you have to compile without OpenMP. See
section `OpenMP' for detail.
- Windows:
See `Building Windows Binaries' to compile.
Data Format
===========
The data format of LIBFFM is:
<label> <field1>:<index1>:<value1> <field2>:<index2>:<value2> ...
.
.
.
`field' and `index' should be non-negative integers. See an example
`bigdata.tr.txt.'
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
-v <fold>: set the number of folds for cross-validation
--quiet: quiet model (no output)
--no-norm: disable instance-wise normalization
--no-rand: disable random update
--on-disk: perform on-disk training (a temporary file <training_set_file>.bin will be generated)
--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.
By default, our algorithm randomly select an instance for update in each inner iteration. On some datasets you may
want to do update in the original order. You can do it by using `--no-rand' together with `-s 1.'
If you do not have enough memory, then you can use `--on-disk' to do disk-level training. Two restrictions when you
use this mode:
1. So far we do not allow random update in the mode, so please use
`--no-rand' if you want to do on-disk training.
2. Cross-validation in this mode is not yet supported.
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
========
> ffm-train bigdata.tr.txt model
train a model using the default parameters
> ffm-train -l 0.001 -k 16 -t 30 -r 0.05 -s 4 bigdata.tr.txt model
train a model using the following parameters:
regularization cost = 0.001
latent factors = 16
iterations = 30
learning rate = 0.05
threads = 4
> ffm-train -p bigdata.te.txt bigdata.tr.txt model
use bigdata.te.txt as validation set
> ffm-train -v 5 bigdata.tr.txt
do five fold cross validation
> ffm-train --quiet bigdata.tr.txt
do not print message to screen
> ffm-predict bigdata.te.txt model output
do prediction
> ffm-train --no-rand --on-disk bigdata.tr.txt
perform on-disk training
> ffm-train -p bigdata.te.txt -t 100 --auto-stop bigdata.tr.txt
use auto-stop to stop at the best iteration according to validation loss
Library Usage
=============
These structures and functions are declared in the header file `ffm.h.' You need to #include `ffm.h' in your C/C++
source files and link your program with `ffm.cpp.' You can see `ffm-train.cpp' and `ffm-predict.cpp' for examples
showing how to use them.
There are four public data structures in LIBFFM.
- struct ffm_node
{
ffm_int f; // field index
ffm_int j; // column index
ffm_float v; // value
};
Each `ffm_node' represents a non-zero element in a sparse matrix.
- struct ffm_problem
{
ffm_int n; // number of features
ffm_int l; // number of instances
ffm_int m; // number of fields
ffm_node *X; // non-zero elements
ffm_long *P; // row pointers
ffm_float *Y; // labels
};
- struct ffm_parameter
{
ffm_float eta;
ffm_float lambda;
ffm_int nr_iters;
ffm_int k;
ffm_int nr_threads;
bool quiet;
bool normalization;
bool random;
bool auto_stop;
};
`ffm_parameter' represents the parameters used for training. The meaning of
each variable is:
variable meaning default
============================================================
eta learning rate 0.1
lambda regularization cost 0
nr_iters number of iterations 15
k number of latent factors 4
nr_threads number of threads used 1
quiet no outputs to stdout false
normalization instance-wise normalization false
random randomly select instance in SG true
auto_stop auto stop at the best iteration false
To obtain a parameter object with default values, use the function
`ffm_get_default_param.'
- struct ffm_model
{
ffm_int n; // number of features
ffm_int m; // number of fields
ffm_int k; // number of latent factors
ffm_float *W; // store model values
bool normalization; // do instance-wise normalization
};
Functions available in LIBFFM include:
- ffm_parameter ffm_get_default_param();
Get default parameters.
- ffm_int ffm_save_model(struct ffm_model const *model, char const *path);
Save a model. It returns 0 on sucess and 1 on failure.
- struct ffm_model* ffm_load_model(char const *path);
Load a model. If the model could not be loaded, a nullptr is returned.
- void ffm_destroy_model(struct ffm_model **model);
Destroy a model.
- struct ffm_model* ffm_train(struct ffm_problem const *prob, ffm_parameter param);
Train a model.
- struct ffm_model* ffm_train_with_validation(struct ffm_problem const *Tr, struct ffm_problem const *Va, ffm_parameter param);
Train a model with training set `Tr' and validation set `Va.' The logloss of the validation set is printed at each
iteration.
- ffm_float ffm_cross_validation(struct ffm_problem const *prob, ffm_int nr_folds, ffm_parameter param);
Do cross validation with `nr_folds' folds.
- ffm_float ffm_predict(ffm_node *begin, ffm_node *end, ffm_model *model);
Do prediction. `begin' and `end' are pointers to specify the beginning and ending position of the instance to be
predicted.
OpenMP
======
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 always run `make clean all' if these flags are changed.
Building Windows Binaries
=========================
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
Contributors
============
Yuchin Juan, Wei-Sheng Chin, and Yong Zhuang
For questions, comments, feature requests, or bug report, please send your email to
Yuchin (guestwalk@gmail.com)