BPR FMRecommender

[parklize]

Hi @ibayer, I'd like to load X_train separately when it is to big to fit at a time, i.e., instead of

fm = bpr.FMRecommender(n_iter=10,
               init_stdev=0.01, l2_reg_w=.5, l2_reg_V=.5, rank=10,
               step_size=.002, random_state=11)
fm.fit(X_train, compares)

I'd like to do something like this..., can I get some advice???

 for i in range(10):
    compares = sklearn.utils.shuffle(compares)
    for [some_parts] in compares:
         fm.fit(X_train[some_parts], compares[some_parts])

where instead of creating a csc_matrix X_train, parts of them are created and loaded for fitting fm.

[ibayer]

Hi the 'n_more_iter` option allows to train the model over junks of data see http://ibayer.github.io/fastFM/guide.html#learning-curves, you just have to swap the training set at every iteration of the loop.

fm.fit(X_train, y_train, n_more_iter=step_size)

[parklize]

Hi thank you for your prompt reply. The option 'n_more_iter` optionIt's only available for FMRegression, but not for FMRecommender, isn't it?

class fastFM.als.FMRegression has fit(X_train, y_train, n_more_iter=0) while 
class fastFM.bpr.FMRecommender only has fit(X, pairs)

from http://ibayer.github.io/fastFM/api.html#fastFM.bpr.FMRecommender doc

[parklize]

Hi @ibayer, if I understand correctly, there should be a method with the parameter 'n_more_iter`, and I tried to modify the fit() function but couldn't make it. I think the ffm_sgd_bpr_fit() in ffm.c should be correspondingly modified. Can you have a look at this ? I'm not familiar with c language. Thanks a lot....

def fit(self, X, pairs, n_more_iter=0):
    """ Fit model with specified loss.

    Parameters
    ----------
    X : scipy.sparse.csc_matrix, (n_samples, n_features)

    y : float | ndarray, shape = (n_compares, 2)
            Each row `i` defines a pair of samples such that
            the first returns a high value then the second
            FM(X[i,0]) > FM(X[i, 1]).
    """
    # The sgd solver expects a transposed design matrix in column major
    # order (csc_matrix).
    X = X.T  # creates a copy
    X = check_array(X, accept_sparse="csc", dtype=np.float64)
    assert_all_finite(pairs)

    pairs = pairs.astype(np.float64)

    # check that pairs contain no real values
    assert_array_equal(pairs, pairs.astype(np.int32))
    assert pairs.max() <= X.shape[1]
    assert pairs.min() >= 0
                    
    self.n_iter = self.n_iter + n_more_iter

    if n_more_iter > 0:
        print 'warm start'
        _check_warm_start(self, X.T)
        self.warm_start = True
            
    self.w0_, self.w_, self.V_ = ffm.ffm_fit_sgd_bpr(self, X, pairs)
    
    if self.iter_count != 0:
        self.iter_count = self.iter_count + n_more_iter
    else:
        self.iter_count = self.n_iter

    # reset to default setting
    self.warm_start = False
    
    return self

[ibayer]

@parklize your right warm start is currently not implemented for SGD (and BPR). I'm not sure, but it might be possible to add warm start support without changing the C code. Have a look at:

https://github.com/ibayer/fastFM/pull/75 https://github.com/ibayer/fastFM/issues/74

[jerry-rubikloud]

@ibayer Is it possible to change the step_size? SGD with a constant step size does not converge well in general or this is a particular implementation of SGD that takes care of that internally?

[ibayer]

@jerry-rubikloud The step_size is independent of the iteration number. It's indeed possible that SGD only converges with a very slow constant step_size in some situations which would come with very slow convergence. Using multiple fit calls with decreasing step_sizes would be helpful but warm_start is currently not implemented for sgd.

A new BPR / SGD implementation is on it's way to solve this issues but it might take a while till we have it ready for release.

[jerry-rubikloud]

A new BPR / SGD implementation is on it's way to solve this issues but it might take a while till we have it ready for release.

@ibayer Thanks for the quick reply and thanks for creating this package. The BPR implementation is actually the reason why I choose this library over the others. :)