fm.predict gives nan values

[chrisbangun]

Hi,

I follow the step-by-step provided in the test_ranking.py in order to use fastFM for ranking problem. Using my own dataset, the model returns nan prediction value.

fm = bpr.FMRecommender(n_iter=2000,
                           init_stdev=0.01, l2_reg_w=.5, l2_reg_V=.5, rank=2,
                           step_size=.002, random_state=11)

fm.fit(x_train_sp_matrix,compares)

y_pred = fm.predict(X_test)

the X_test is simply the copy of my training data. what would be the case where fm model returns nan? any help would really be appreciated. thanks

[todor-markov]

I have the same problem but with sgd.FMClassification

[yushcs]

Is there anyone know how to solve it?

[ibayer]

Unfortunately many issues can lead to nan predictions.

• Make sure you model inputs are sound.
• If the predictions are nan check if the model parameter are nan too.
• Both bpr.FMRecommender and sgd.FMClassification use stochastic gradient based solvers for the parameter estimation (fit). This means they are sensitive to the step_size and initial values init_stdev hyper-parameter. Getting this values wrong can often lead to 'nan' predictions.
• The bpr implementation is not very robust and should currently only be used for small scale experiments.

@todor-markov I would recomment to use use the als/mcmc solver instead.