Why not use all articles when testing?

[01joy]

Hi,

Your 3D-CNN is a very good idea when recommending news, and I'm adapting your model to my data.

I have one question after reading your codes. In your test function, for each user, the user_in vector is the same as it in train function, and the article_in contains 99 random articles and 1 target article, see code link. Why not use all articles when testing?

For example, for a user, she reads [article1, article2, ..., article10, article11, article12]，when training, the user_in=[article1, article2, ..., article10], and the positive article_in=article11. When testing, user_in is the same as when training=[article1, article2, ..., article10], and article_in contains randomly sampled 99 articles, and the target article 12. In your case, the model is more likely to predict the correct article12. However, if the article_in contains all articles, including the article11 which was the label when training, the model is definitely to predict the article11, because it is the label when training.

So, Why not use all articles when testing? In the real world scenario, all articles should be the candidates when recommending.

I'd appreciate it if you could help resolve this issue. Thank you very much.