solution for cold-start problem?

[joshbedo]

I was wondering if their is anything planned or done to solve the cold-start problem. The scenario is you haven't taken any actions so no recommendations are returned but since no recommendations are returned you can't take action on anything.

[grahamjenson]

No work inside of hapiger. My recommendation is to populate the dataset with fake data, like what a person who is very typical would do. Ensure that every thing has at least one action against it, then you should be able to generate recommendations for those items and people who are new.

[mgreer]

What I do is this:

1. I have an anonymous user with id=-1, so all anonymous users share that hapiger id when watching, sharing, looking for videos
2. When someone gets their own id (signs up), I backfill their recommendations with the anonymous ones for a while, since those are usually the "starter" ones. This gives them a bit of a start, since by selecting from among these they start building a set of particular interests. If they comment or share, even more so. I drop the crutch after they watch 5 videos.

Oh, I'm also caching those -1 recommendations, so the extra call is very fast.

[ttopholm]

@grahamjenson just a dumb question, if I choose your approach, by giving a action "view" on all items, but no recommendations, will it still give me recommendations.

[grahamjenson]

It should do, given it would match you with really low ranking with everyone else who also viewed those items.