A recommendation engine needs information about what posts you are glad to have read and ideally what posts you read but did not fund useful. So if the engine knows for each user (1) the set of posts they’ve read [1] and (2) the set of posts that they’ve voted up, then we have an evaluation criterion: did we choose to show people posts they voted up?
You’d then need to figure out features and write code for them so the learning algorithm could find user correlations. Set up a svm or something to get probability of upvoting given viewing. Then you use some sort of multi armed bandit algorithm so you continue to gather information.
[1] This isn’t perfect, because you can open a post without reading it. We could detect scrolling and log how much of it they actually read, but people might not like the privacy implications.
Better than recency, perhaps the top scoring posts of all time?
A recommendation engine needs information about what posts you are glad to have read and ideally what posts you read but did not fund useful. So if the engine knows for each user (1) the set of posts they’ve read [1] and (2) the set of posts that they’ve voted up, then we have an evaluation criterion: did we choose to show people posts they voted up?
You’d then need to figure out features and write code for them so the learning algorithm could find user correlations. Set up a svm or something to get probability of upvoting given viewing. Then you use some sort of multi armed bandit algorithm so you continue to gather information.
[1] This isn’t perfect, because you can open a post without reading it. We could detect scrolling and log how much of it they actually read, but people might not like the privacy implications.