Page MenuHomePhabricator

Add weight parameter to morelikethis CirrusSearch feature
Closed, DeclinedPublic

Description

In GrowthExperiments we are using morelikethis as an approximation of topic search. It would be nice to be able to do (...) morelikethis:A|B morelikethis:X|Y and get a result set which is a mix of articles similar to A or B and articles similar to X or Y. What often ends up happening is that one of the morelikethis article sets ends up more "powerful" then the other, so all results are via A/B (or the other way around).

One way to address this would be to allow weighted morelike queries: we'd pre-calculate the top scores for (...) morelikethis:A|B and (...) morelikethis:X|Y (the (...) can only take a limited number of values) and use them as inverse weights, with something like (...) morelikethis:A^0.7|B^0.7 morelikethis:X^0.3|Y^0.3.

Event Timeline

Tgr created this task.Jan 11 2020, 12:56 AM
Restricted Application added a subscriber: Aklapper. · View Herald TranscriptJan 11 2020, 12:56 AM
EBernhardson triaged this task as Medium priority.Jan 13 2020, 7:36 PM
EBernhardson moved this task from needs triage to Current work on the Discovery-Search board.
Tgr added a comment.EditedJan 21 2020, 11:00 PM

To clarify, for Growth this is a nice-to-have since if all goes well soon we'll be using ORES for topic search. But it seems to me like a sensible thing in general (of course often a sensible thing for the user is feature creep for the maintainer).

Note to self: we'll probably need the equivalent functionality for ORES search.

EBernhardson added a subscriber: EBernhardson.

Moving off workboard, with the current articletopic implementation replaceing morelikethis I don't think this is necessary

Tgr closed this task as Declined.Wed, Mar 25, 6:55 PM

Agreed. We can reopen if there's a use case.