Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

IMO, you could highlight the differences between similar titles.

It's easy for recommendations to get stuck on a local maxima if they only look at one metric at a time, like "similarity" weights. But if you have a lot of metadata about each title, you can break out of those "loops" by sprinkling in metrics like ratings/genres/release date/popularity/etc. This doesn't have to hurt from a performance perspective, either; you can filter on the same single metric, but request more recommendations than you need and pluck out a pseudo-random set in the application logic.

That also lets you provide context for the recommendation. "It's like this, but [older/more obscure/with vampires]."



Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: