recommenders-team/recommenders
Voir sur GitHub[ASK] [DISCUSSION] What kind of news is recommended
Open
#1 251 ouverte le 26 nov. 2020
help wanted
Métriques du dépôt
- Stars
- (17 706 stars)
- Métriques de merge PR
- (Merge moyen 6j 16h) (10 PRs mergées en 30 j)
Description
Scorer and recommendation items
if you look in the function run_fast_eval what exactly we recommend and add the following changes:
for (
impr_index,
news_index,
user_index,
label,
) in tqdm(self.test_iterator.load_impression_from_file(behaviors_file)):
pred = np.dot(
**np.stack(news_vecs.values(), axis=0),**
user_vecs[impr_index],
)
**pred = np.argsort(pred)[::-1]**
it will become clear that at the beginning of the list of recommendations there will ALWAYS be those articles, the number of which on the train as 1 was greater. which means, in fact, not quite relevant items are recommended. and not even the most popular, since this is only the number in the target, and not in all sessions
similarly, if you shoot a metric in a slice not an impression but in a user slice, that is, for 1 user and several 1 targets, then the metric will immediately drop by more than 2 times