simple-libfm-example-part1

原文：https://thierrysilbermann.wordpress.com/2015/02/11/simple-libfm-example-part1/

I often get email of people asking me how to run libFM and having trouble to understand the whole pipeline. If you are verse in Machine Learning, a first step is to take a look at Steffen Rendle’s paper ‘Factorization Machines‘ and this one too:Factorization Machines with libFM

I’ll try to explain how to train different kinds of models with the 4 different learning algorithms that libFM provides and use the features of libFM (like grouping and relations)

But first, here is a toy example of how each file should be. (Was posted in the libfm google group)

Simple example for 2 users and 3 items. We have 2 users, 3 items in our training set and now you want to test on the same 2 users, but now you have 4 items (the same 3 from training + one new))
Each user has a categorical feature age [“18-25”, “26-40”, “40-60”] and each item has a numerical feature price.

I one-hot encoded the users:

0 is User1
1 is User2

Same thing for items,

2 is Item1,
3 is Item2,
4 is Item3,
5 is Item4

The categorical feature age need to be one encoded too

6 is the category “18-25”,
7 is the category “26-40”,
8 is the category “40-60”

And finally the numerical feature price for the item

9 will represent the price feature

One sample can be:

5 0:1 3:1 6:1 9:20
#User1 who is 23yo is giving a rating of 5 on Item2 which costs 20 euros

We can then construct example and create a training and test set.

train.libfm

5 0:1 2:1 6:1 9:12.5
5 0:1 3:1 6:1 9:20  
4 0:1 4:1 6:1 9:78 
1 1:1 2:1 8:1 9:12.5
1 1:1 3:1 8:1 9:20

num_features = 10  #Computed on the highest integer value that represents a feature (here 9 for the Item price) + 1 (because we expect people to start at 0)

test.libfm

0 1:1 4:1 8:1 9:78
0 0:1 5:1 6:1

num_features = 10 #Computed on the highest integer value that represents a feature (here 9 for the Item price) + 1 (because we expect people to start at 0)

For the test, I have two samples I want prediction. The 0 doesn’t really have any effect in testing (Only useful if you have the true value, then libFM will output the RMSE error on it but will not use it to train the model)
Just to be sure, here is the meaning of those two samples in test:

0 1:1 4:1 8:1 9:78
#Here User2 who is 41yo is rating Item3 which costs 78 euros and we gave a rating of 0 because we don’t know yet the real rating

0 0:1 5:1 6:1
#We want to know which rating User1 who is 23yo will give to a not-yet seen Item4 and we don’t know the price

This format is the same as for libSVM

From here you have two files: train.libfm and test.libfm (the extension doesn’t matter)

You can then run libFM like this for regression (predicting ratings):

./libfm -task r -method mcmc -train train.libfm -test test.libfm -iter 10 -dim ‘1,1,2’ -out output.libfm

So the model was train using [MCMC (-method mcmc)] on [10 (-iter 10)] iterations using a [linear model (+bias) and using factorization with 2 latent factors. (-dim ‘1,1,2’)]

You will then get some output out of the command line and prediction will be written in the file ‘output.libfm’

Discussions
This is of course a toy example but show you what you can use in libFM to train your model.
I wouldn’t recommand using the price feature like this but maybe do some transformation like log to avoid having a feature with large value but I hope you get the point.