search

wuch15 / KDD-NPA

Resources for the paper "NPA: News Recommendation with Personalized Attention"
47 stars 12 forks source link

user encoder #4

Open shainaraza opened 4 years ago

shainaraza commented 4 years ago

Should I consider this as user encoder according to NPA model = Model(candidates+all_news_input+[user_id], logits)

wuch15 commented 4 years ago

This module actually includes both user encoder and click predictor. The output is the predicted click scores of "candidates".

shainaraza commented 4 years ago

thanks

shainaraza commented 4 years ago

and you are considering titles only , not the body of the news,

shainaraza commented 4 years ago

ideally, if I have to incorporate GRUs like in LSTUR, where should it be? thanks in advance

wuch15 commented 4 years ago

and you are considering titles only , not the body of the news,

In this work, we only consider news titles, and news bodies are incorporated by our another work NAML. But you can simply concatenate news title and body in NPA.

ideally, if I have to incorporate GRUs like in LSTUR, where should it be?

I think the method in LSTUR is not compatible with personalized attention. You may need to output each state in the GRU network first, and then apply personalized attention to the hidden state sequence.

shainaraza commented 4 years ago

thanks you very much