Comment by pitherandd | Hacker News Reader

Comment by pitherandd | Hacker News Reader

Parent Full thread

Another fantastic reply! I really need to get this to take off already so I can hire you before you make your own and out-compete me.

The more I read this reply, the more I agree with it, and I think it may actually not be that difficult of a change. Card sets could be integrated into the existing cluster concept and I could just give users the ability to choose which sets (clusters) they swipe on and the weighting that they apply to each. They could also decide which clusters should be factored into their similarity matching. I _think_ this will all work with the existing CUBE concept, which is exciting, because many other proposed solutions by others didn't fit nicely within that mathematical structure.

You've honestly given me a lot to think about and I think I see a better way forward now. Your insight really increased my mood because I think you've discovered something very important that I am likely going to be spending quite a bit of time on in the next coming weeks and months.

Thanks so much!

I would advise against folding it into clusters, and instead have each cardset be its own thing. What you have right now is a cardset that I'd label "General" -- abstract the backend so you can create different sets of cards ("Food", "Fashion", etc). If manually curate cardsets, you won't even need to worry about clustering. Yes, it's a shame you spent time on it, but no clustering will ever beat manually creating sets of cards.

In terms of producing a "total match score" with a user, you compute a match score for each cardset that both users have, then use a simple normalized linear combination to get the total.

If users A and B have cardsets X Y and Z in common, you would produce similarity scores "S" for S(A,B,X), S(A,B,Y), and S(A,B,Z). Then, you use the weights that user A selects for each cardset (W(A), W(B), W(C)), normalize such that they add up to 1 but maintain their ratios, and compute total similarity of A matched to B as: W(A) * S(A) + W(B) * S(B) + W(C) * S(C).

As long as you have pre-computed the scores between all user-cardset pairs (your scaling pain point), computing match scores even with weights is trivial and fast.

> You've honestly given me a lot to think about and I think I see a better way forward now. Your insight really increased my mood because I think you've discovered something very important that I am likely going to be spending quite a bit of time on in the next coming weeks and months.

Happy to hear. I've been working on my own project for close to a year and am close to launch, so I think I understand where you are coming from.

It is practically impossible to view your product as someone unfamiliar with it would. So, that leaves you asking for feedback. Next, it is really difficult to distill user feedback (such as found on this thread) into things you should actually work on. Is a comment just a vocal minority complaining or an indication that some concept should be changed? I think you're doing a good job taking feedback to heart and I'm really rooting for you.

> Another fantastic reply! I really need to get this to take off already so I can hire you before you make your own and out-compete me.

If/when it takes off, just make me an adviser and shoot a couple percentage points my way!

Reply on news.ycombinator.com