MatrixFactorization predicts the same values for all testing instances

[RF]

The MatrixFactorization is always returning the same rating value for all testing instances. This does not occur with any other algorithm that I have used.

The ratings were normalized to [0.0,1.0] and the command line is:

rating_prediction --prediction-file=<FILE> --prediction-line={0},{1},{2},0 --training-file=<FILE> --test-file=<FILE> --recommender=MatrixFactorization --recommender-options="num_factors=44 regularization=0.29 learn_rate=0.087 num_iter=60"

I'm using MyMediaLite 3.10.

[Zeno Gantner]

Hello,

How does your data look like?

Is there an overlap between training and test file in terms of user or item IDs?

Cheers,

   Z.

--
You received this message because you are subscribed to the Google Groups "MyMediaLite" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mymedialite...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[RF]

Hi,

there is no overlap between training and testing.

[Zeno Gantner]

You mean every user and item in the test set do not occur in the training set?

If this is so, there is nothing that the model could learn. But other models neither. And the prediction would be the same (the global bias) for any of those users and items.

Cheers,
   Z.

[Reinaldo Silva Fortes]

No, there are training data for the majority of users and items, the problem occurs only for matrix factorization.

--
You received this message because you are subscribed to a topic in the Google Groups "MyMediaLite" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/mymedialite/f8M_JReXG_k/unsubscribe.
To unsubscribe from this group and all its topics, send an email to mymedialite...@googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

--

-----------------------------
Reinaldo Silva Fortes

Universidade Federal de Ouro Preto - UFOP

Instituto de Ciências Exatas e Biológicas - ICEB

Departamento de Computação - DECOM

Sala 17, terceiro andar do ICEB-III

Página Web: http://www.decom.ufop.br/reinaldo/

E-mail institucional: reif...@iceb.ufop.br

Telefone: (31) 3559-1603

"O pessimista queixa-se do vento, 
o otimista espera que ele mude e 
o realista ajusta as velas." 
(Willian George Ward)

[Zeno Gantner]

Which other methods did you try? What kind of results did you get?

Z.

[RF]

I have attached a file showing a small sample of results. You can see what are the tested methods and the normalized rating values.

The Matrix Factorization not always result in the same value, but showed a coefficient of variation very close to zero.

The database is very sparse, having 77,805 users, 185,973 items, but only 433,671 ratings.

[Zeno Gantner]

Sorry for the late follow-up -- if you are still interested in this: What kind of RMSE do you see for the different models?

The dataset is indeed very sparse.

I wonder if the results are better for certain models ...

Cheers,

   Z.

--

You received this message because you are subscribed to the Google Groups "MyMediaLite" group.

To unsubscribe from this group and stop receiving emails from it, send an email to mymedialite+unsubscribe@googlegroups.com.