Unable to reproduce the results on Criteo.

[cjxnn]

We are trying the author's PIN code on Criteo and Avazu. We are able to reproduce the AUC score of 78.72% on Avazu. But we can only achieve an AUC score of 80.18% on Criteo. However, if we use a different embedding size for each field, we are able to get an AUC score of 80.21% on Criteo. But this is not the setting claimed the paper. Could the authors clarify on this issue?

[Atomu2014]

Hi,

I think you can uncomment line 285 and 574 of tf_main.py, and tune "eval_level" and "decay", e.g., eval_level = 5, decay = 0.8.

Learning rate decay could alleviate overfitting, but the code is commented due to incompatibility with the distributed version. I used learning rate decay on every model on Criteo, which showed improvements. I think I forget to claim the special setting of Criteo in the paper.

Besides, table 6 compares models under restricted settings. You can find better results of corresponding models in table 9, which are carefully tuned with grid search.

Willing to provide further help~

[Atomu2014]

Be careful that learning rate decay is not fully implemented in the distributed version.