Giters

gabrielspmoreira / chameleon_recsys

Source code of CHAMELEON - A Deep Learning Meta-Architecture for News Recommender Systems

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Why not use the author info during the ACR?

zippeurfou opened this issue · comments

commented

Hi,
I have read the paper and I am wondering why the author was only considered during the NER part and not the ACR part. The paper doesn't really explain why this could not be considered as a metadata attribute or at least i am missing it.
Best,

commented

Hi,
In the ACR module, the Article Content Embeddings are trained in a side task: classify articles' categorical metadata, such as Category or Tags, based on the article title and text. You could also use the Author as another target categorical attribute for training. Therefore, I am afraid that content embeddings trained to predict who is the article author might not lead to good content representations for recommendation, compared to categories, which are defined by Editors and are in general very related to the article content.
Is that clearer?

commented

Thanks @gabrielspmoreira for you answer.
I was more thinking of using it as a metadata input feature for ACR training.
For example you do not use it for the addressa ACR training but it get passed to the NAR. On the other hand it seems to be used for the gcom ACR training. I was wondering why it was like that?