scores are glicko-2 scores, with a rotating 10-rating window for both user tag and link tag scores, but a decay window of a year for link tag scores and a decay window of a month (30 days) for user tag scores
decays are checked lazily; that is, they are only performed when a score is requested (such as when something is liked, disliked, or when a feed is updated and the score needs to be checked for a link
feeds are determined with the following algorithm:
-
select every link that is both unseen by the user and has a tag selected for inclusion by the user
-
for each link, intersect the tags of the user and the link and
i. sum the scores of all of the user's tags and convert each tag's score to a percentage of the sum
ii. take the link's tags and multiply each of them by their corresponding tag percentage calculated from the user's tag
iii. average the resultant scores, creating the "relative overall score" of the link
-
sort the links by their overall score and divide the links into four segments, and pick randomly, starting from the top, 4 links, 3 links, 2 links, and 1 link. this is the user's feed
just run the avg function over all of the fields
bucket by value, sort (from top) by decreasing confidence
so, if a player had a rating of 1500, dev of 350, and a volatility of .6, and another had a rating of 1500, a dev of 600, and a volatility of .6, the first player would be placed above the second
you sort starting with the rating (which rating is higher), then go to the deviation (which deviation is smaller), and then go to volatility (which volatility is smaller)
the reasoning is that smaller values for the latter two parameters indicate greater confidence in the first value, and intuition would suggest that a score with greater confidence should be placed above scores of lesser confidence if the rating is the same