NaN outputs

[yangji9181]

Hi Ben,

It feels like the current code is still suffering this same issue (#19 (comment)):

when I tried to run the 'basic usage' on a few small 100*200 random 0-1 matrixes, the output scores are always nan. Could you try to look into that again?

Thanks.

[yangji9181]

Here is the exact code I have tried. I am not sure what is causing the problem, and my purpose here is just to make sure your framework works before I deploy it into my experiments.

X=np.random.randint(5, size=(100,200))
X0=coo_matrix(X, dtype=np.float64)
model = implicit.als.AlternatingLeastSquares(factors=50)
model.fit(X0)
print(model.user_factors)
print(model.item_factors)

Thanks!

[benfred]

Hmm - this works for me (on both OSX and Ubuntu). What version of implicit are you running? Also what version of python/numpy etc are you running ?

In [1]: import numpy as np

In [2]: from scipy.sparse import coo_matrix

In [3]: import implicit.als

In [4]: X=np.random.randint(5, size=(100,200))
   ...: X0=coo_matrix(X, dtype=np.float64)
   ...: model = implicit.als.AlternatingLeastSquares(factors=50)
   ...: model.fit(X0)
   ...: print(model.user_factors)
   ...: print(model.item_factors)
   ...: 
[[-0.50830597  1.77890754  3.72887993 ...,  5.42687702  0.73224628
   2.95396447]
 [ 4.16104603  3.23049259  1.36413717 ...,  2.49195838  1.41964936
   1.75384963]
 [ 1.88410449  1.15888965  1.82405496 ...,  2.37375712  2.4774456
   0.28280351]
 ..., 
 [ 2.13711357  0.86102092 -0.43549636 ...,  2.66490889  4.2622385
   3.61764646]
 [ 3.21060467  1.86145854  3.29952288 ...,  2.49085307  2.59494758
   0.53273898]
 [ 1.63236988  3.20456839  2.73510027 ...,  2.82707405  2.82944465
   1.39465821]]
[[-0.0089963   0.09548787 -0.03623041 ...,  0.01521569  0.02308306
   0.05071232]
 [-0.00933617 -0.00377901  0.04466964 ...,  0.06146013 -0.00124603
   0.01044547]
 [ 0.00631106  0.01320372  0.03328578 ...,  0.00114081  0.01793537
  -0.07993541]
 ..., 
 [-0.08164594 -0.07154793  0.06172384 ..., -0.02570276  0.06908227
  -0.02065015]
 [-0.00280522  0.04038418  0.05980453 ...,  0.07462172  0.01221859
   0.01574233]
 [-0.00381051 -0.00444142 -0.00367571 ..., -0.00306982 -0.02002124
   0.01943841]]

In [5]: np.isnan(model.user_factors).any()
Out[5]: False

In [6]: np.isnan(model.item_factors).any()
Out[6]: False

[yangji9181]

Hi Ben,

Thanks for replying. I was using miniconda2, but after I switched back to Mac's default python, it worked fine.

[yangji9181]

Hi Ben,

By the way, I noticed that the model is prone to overfitting. Besides setting larger regularization term, is there a way to do something like adding a mask on the training data, so that the recovery loss is only computed on non-zero values? I think this might help reducing overfitting too.

Thanks.

[benfred]

For the NaN issue - my best guess is that you still have an older version installed with your miniconda2 install.

For overfitting you can either increase the regularization parameter, or alternatively you can decrease the number of iterations you train for.

Its not possible to compute the loss function here without the non-zero values - they are the negative examples, and computing the loss with only positive examples would mean that a model that predicted 1 for all items would score perfectly (if that makes sense). If you want to weight the non-zero values higher, you should just increase the confidence in the Ciu matrix.