Recommended is too precise/non-random

I just wanted to comment that the new recommended feature seems to always come back with over half of the same images and has done so since its beginning. And I have close to my gold cap worth of favorites, which is a lot to pull data from, and could be the reason. Anyways, I'm sure I'm not the only one and someone probably agrees with me.
Perhaps there could be a change in a randomness value or better yet, options for what parameters could be changed, such as whether or not to return favorited images, a slider for how far you wanted the program to deviate from your preferences, or even choosing a set of images to find recommended content based off that set alone, similar to how individual posts have recommended sections but for the select set.
Thanks, you guys rock!

Yeah, it seems to freeze at a local maximum at some point where you get to the point that you have to favorite its recommendations simply to get new ones (which would arbitrarily move you in the direction it picks for you), or you can't use it anymore since there's nothing you can pick that you feel strongly enough to favorite. It'd be nice if we could "pass" on recommendations to suppress them, and potentially provide the recommendation system additional information. It might add complexity to the recommend though, and require a modification to the database to give that information somewhere to live. I'm not sure how feasible or infeasible that might be. The post level recommendations continue to be pretty good though and give you greater variety.

Even then, I have favorited most of the suggestions and it does not remove them.

Agreed, on both the "pass" idea and post level recommendations being good.

well here's hoping they'll actually fix it then

Bump since this is still a problem :(

i hope you can fix it!

its a very good feature but i faved 80% of the suggestions and get no new recommendations >-<

hope its ok to bump this since i would love to see that feature working again

bump

it's been broken since you first asked

and is still broken entirely

Still broken...

There have been some recent changes to make this a bit better, but the system is still very experimental and should be considered a beta. The issue with recommendations not changing very often is still present. This is because the system is showing you your top-100 recommendations of all time, which don't change very often. You can use tag searches now to work around this. Searching age:<1month on the /recommended_posts page is a good thing to try. The idea about giving feedback to the system (to mark recommendations as good/bad/ignored) is something that will probably be implemented eventually.

Some tag searches seem to give far fewer results than I'd expect to be recommended - is there a limit to the total number of posts that can be recommended across the site as a whole or something?

For instance, doing a recommended posts search for shiranui_(kantai_collection) gives me just 2 results, and they're not the two posts that I haven't favourited which have either the highest score or the highest fav count, so it isn't being limited in that way, and there are people who have favourited images that are not showing up with whom I share a very large number of favourites, which you would think would make the images in question appear, albeit not necessarily at the top of the list.

Tag searching has known limitations. The system works by generating a certain number of recommendations for you first, then filtering favorites and tags out afterwards. So really you're just searching from a relatively small set of pregenerated recommendations, which may not contain the tag you're looking for.

There's a secret max_recommendations param you can use to force it to generate more recommendations:

https://danbooru.donmai.us/recommended_posts?search[user_name]=evazion&search[max_recommendations]=50000&search[post_tags_match]=shiranui_(kantai_collection)

Currently the max varies by user but is capped at 50,000 recommendations, which is why searching for smaller tags is less likely to work.