Hacker News Reader

Comment by teej | Hacker News Reader

From the "how does this work"

> How do products trend? What causes a product to trend? Products are ranked based on the number of users who have recommended the product in the past month (we’re using a rolling 30 days).

Ok so this isn't trending, it's popularity. Typically you want to normalize such that items with evergreen popularity get sifted out and you're left with items growing in popularity.

A marginally better approach might be to look at % of like growth month over month, with some floor of likes so that going from 1 to 100 likes doesn't dominate. This isn't perfect but it should kick google drive, Dropbox, and git out of the list.

View on HN

We're looking into ways to improve the ranking, so thanks for sharing that teej. We actually started with a normalized and more complex algo that was more along the lines of what you suggest, but the resulting feed had mostly random products and was not very helpful for more casual users. It also made it difficult to understand why a product was ranked. I think you're right that as it stands, we can do more to surface younger products. So far, we're doing it manually with "wildcards".

I mentioned in another comment that one of our bets is that since a user can only recommend a product once, over time the more obvious products should find it relatively more difficult to get new recommendations and newer products will surface.

e.g., if we did a filter of recommendations by YC companies (or more extreme - YC founders), then something like Slack would be popular the first month, but then have few remaining votes. Thoughts?

You need to separate the concept of popular and trending. Make two lists. Have a popular list at a couple fixed time grains - all time, last year, last 30 days (what you have now). Then for your trending list use a z-score like described in this stack overflow post [0] to normalize. I honestly wouldn't invest a ton of time trying to do something fancier then that. You're trying to rank with one signal, you can't expect computers to perform magic.

I'd also consider a third list - highlighted. Have your team internally pick something new and interesting once a month. Finding novel and interesting things is something humans are really good at.

[0] - http://stackoverflow.com/questions/787496/what-is-the-best-w...

Thanks. Great link. It does seem that having multiple feeds is emerging as a consensus suggestion.

Reply on news.ycombinator.com