In the general sense we have an existing `order by` that follows the
following format: `order(x) = f(x) * rand() ^ ( 1 / g(x))`
For the present [20220603-variant-b][1], `g(x)` is the article’s score.
And `f(x)` is 1.
This pull request introduces a few named variations for the above
`order(x)` formula.
The three named formats each use a different `g(x)` function; namely
`greatest(0.1, ln(1 + greatest(0, public_reactions_count)))`. The more
public_reactions the article has the closer to zero the resulting
positive number will be. And raising a random number, with range
`(0..1)`, to the inverse of that number will tend to mean that artciles
with many reactions will tend to group towards the higher end of the
random range, whereas articles with few reactions will have a more even
distribution across the range.
Further the `ln` function helps dampen the weight we give to reactions.
The variation within these three order levers is on the `f(x)` portion:
1. Using `published_at`
2. Using `last_commented_at`
3. Using a mix of `published_at` and `last_commented_at`
In all cases, those dates are converted to an integer. A date 2 weeks
ago will result in a number less than a date 1 week ago which will be
less than a today.
The idea for the mix of `published_at` and `last_commented_at` is to see
a mix of things where some “older” articles (from the underlying result
set) will get a chance to show up earlier in the feed.
My rational for this order structure is because the `RANDOM() ^ (1 /
score)` is showing possible success in a current experiment. And in
past trials long past, it rose to the top as a likely and viable
contender.
These new `order_by` lever patterns introduce three things:
1. A variable based on a date in time
2. A dampening of what can be a large number; (we have scores in the
publication range that are between -200 and 750 so I think it is
important to apply a dampening. For example `(0.1 ^ (1/100)) =
0.977` and `(0.01 ^ (1/100)) = 0.955`; in other words these large
scores aggressively pushed their corresponding posts to the top of
the relevance; even as we weren't explicitly querying for those
scores.
3. With the score being quasi-magical, it doesn’t make sense to sort on
that value but to instead sort on a property that represents what
we’re setting as our experiment goals (e.g. publishing something or
commenting on something and/or reacting to something)
Note: it would be feasible to extract these into a singular order lever
that we would configure. But that's a future concern.
Also note, none of these are yet part of any experiment, but are instead
being "queued up" to add to our available corpus.
Closesforem/forem#17834
Related to:
- forem/forem#17827
- forem/forem#17826
- forem/forem#17833
- forem/forem#16128 :: for past order queries to seek as inspiration
[1]:https://github.com/forem/forem/blob/main/config/feed-variants/20220603-variant-b.json
736079e28c