[ad404b8a372f2b9]

It occurs to me you could switch the curation and trend gathering steps to minimize the amount of work on your end. Instead of going "twitter trend -> filter out the bad ones with in-house experts" you could go "follow experts -> find out what trends among them". That's my usual workflow when using twitter.

[mlent]

that is already what i'm doing pretty much. i accept blog submissions and try to look for signs of quality, and then monitor for mentions of their sites.

as you suggested though, i didn't know the ML space enough to know that Machine Learning Mastery publishes daily and has an army of people who like and retweet ANYTHING they publish.

so it makes me wonder if i need to have some kind of dampening effect or how i can adapt the algorithm to handle that.

[zerkten]

Get some experts in the area to rank your existing content and suggest weightings for their recommended sources. This is the less from launch after launch of really great sites that aggregate information.

Let's take Stack Overflow as an example. Jeff found a small group of experts and expanded it. They seeded both questions and answers. They didn't bring on just one or two experts though, they brought on enough to ensure a good distribution (not perfect) of viewpoints and then reviewed before expansion. They kept repeating this and didn't optimize for just one kind of developer (Django over Java.) All segments of developers tended to need the same features, but it would show up with one segment first. Getting answers on some topics wasn't possible until the product was more mature. Kill crap ruthlessly like SO did with downvoting and moderator-led deletion.

If you are building an ML model then you are going to need to find a range of experts and either seed from what they are sharing, or create a review system. You can reward people with kudos on a contribution page, donations to open source projects or charities (even on behalf of a group of them), or find another way to motivate them. It just needs some hustle, but you've got to forget about purity, be open about how your model works, and iterate.