PPAD-AE, GRU or biGRU?

[israa-alqassem]

Hi,

Thanks for implementing and publishing this interesting method. I saw the "cell" type parameter is set to GRU in the file hp_df.csv, however, in Fig. 3 in the paper you showed that you used BiGRUs with encoder-decoder architecture. I wonder whether you used BiGRU or GRU for the results reported in the paper? You mentioned that you used the decoder to generate representations up to n visits, in that context, why it's useful to have a bidirectional model since the data is generated (not real data)?

I appreciate your valuable input.

Thanks, Israa

[Mohammad-AlOlaimat]

Dear Israa, Thank you for your comment! The cell in this architecture is a hyperparameter, and could be GRU, LSTM, BiGRU, or BiLSTM. Although in Fig. 3 we presented BiGRU cell, some for the results reported in the paper have been generated using cell other than BiGRU. I would suggest that you take a look into the supplemental document of PPAD paper.

Thanks, Mohammad

[Mohammad-AlOlaimat]

Regarding to 'You mentioned that you used the decoder to generate representations up to n visits, in that context, why it's useful to have a bidirectional model since the data is generated (not real data)?' The generated representations up to n visits are generated in a casual or autoregressive manner, so a bidirectional model could be used. I hope that I answered all your concerns.