Add Ranking Example about fastfm

[ibayer]

Ok,
fastFM uses the Bayesian Personalized Ranking [0] (BPR) loss with stochastic gradient descent (SGD) for ranking. Usually for SGD you calculate the gradient for each sample and then update the parameter; for a pairwise loss the gradient is calculated for a pair of samples.

Let's assume that we have only 1 query and 3 items. Each user / item can be represented by a vector x, and we construct an example by concatenating x = [x_query, x_item] query and item vectors.
We further know that item1 is the only correct answer, we therefore want the model returns the highest value for x = [x_query1, x_item1].

training

fastFM needs two input files for training:
The first lists all samples (see: https://github.com/ibayer/fastFM-core/blob/master/demo/data/train_ranking)

    query1, item1
    query1, item2
    query1, item3

The second lists all possible pairs (i, j) such that sample_i > sample_j. This would
be

   0 1
   0 2

in our example (see: https://github.com/ibayer/fastFM-core/blob/master/demo/data/train_pairs)

test / prediction

We have learned a factorization machine with BPR that behaves like a scoring function.
We can now use this model like a regression model (per user) except that we use the prediction to sort the
samples. Back to our example: if we would like to know if item2 is better then item3 for query1 we construct the following input (X_test):

    query1, item2
    query1, item3

and evaluate the model

    y_pred = fm.predict(X_test)

We can now check if y_pred[0] > y_pred[1] to answer the question if item2 is better then item3 for query1.

Does this make sense?

• [0] http://arxiv.org/pdf/1205.2618.pdf

[ibayer]

It's more a matter of time. But you are right it's a great idea to add a ranking example to the tutorial (http://ibayer.github.io/fastFM/index.html).

In the mean time you could have a look at https://github.com/ibayer/fastFM/blob/master/fastFM/tests/test_ranking.py .

[BrianMiner]

An example would be helpful. I am not sure how to structure the X and y. I have seen where each record is a (e.g. features for a query document pair) along with a single relevance rating and then there is an external file or an ID on the record that tells how to group the records into (e.g. queries). That doesnt seem to be the case here though,

[wzhd]

Thanks for your explanation. However, I'm confused by the first input file train_ranking.

0:6 1:1
0:2 1:3
0:3
0:6 1:1
0:4 1:5

The third line appears different.

[ibayer]

You can read the third line as: query with value of 3 for feature 0 and item with only zero valued features.
I agree that having an item with only zero valued features is a bit confusing.