You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Our home feed in Atlas has arguably gotten worse from when I originally raise this issue: #245
I believe we currently have
introduced higly reliable language filtering
allowed all videos from all categories
tried to improve bad channel weights set prior
although this latter point may not have been executed succesfully. We are also going to be filtering out shorts soon, but they also continue to negatively affect the feed.
Currently, the feed looks worse than ever, and there is substantial operational problems in understanding the ranking algorithm at any time, in particular how it trades of recency and quality, and also even what the true quality of a given channel really is.
Proposal
As a sort of short term emergency, I propose we introduce a fully manually curated public home feed.
Data Model
We add a new field include_in_home_feed: boolean to videos, make it false by default.
Query
We introduce a new query called dumb_public_feed which randomly selects up to STATIC_HOME_FEED_LENGTH videos so long as
their language is suitable
include_in_home_feed is true
not censored
and returns them in random order, so expected to be different from query to query, even if data is same. I believe in PostgreSQL this query would look something like this
SELECT*FROM videos ORDER BY RANDOM() LIMIT STATIC_HOME_FEED_LENGTH;
Operator API
We add two operator endpoints
list all videos with include_in_home_feed
set include_in_home_feed for video with given ID.
Operationally
The curators should try to update overall composition of videos over time, the scope and frequency of
Costs
This does not take recency into account, which I think is fine for now
The time required to update the pool of videos is unknown, and could be large, but again, the name of the game here is improve content selection, so if we have to accept some older content, thats worth it.
The text was updated successfully, but these errors were encountered:
Background
Our home feed in Atlas has arguably gotten worse from when I originally raise this issue: #245
I believe we currently have
although this latter point may not have been executed succesfully. We are also going to be filtering out shorts soon, but they also continue to negatively affect the feed.
Currently, the feed looks worse than ever, and there is substantial operational problems in understanding the ranking algorithm at any time, in particular how it trades of recency and quality, and also even what the true quality of a given channel really is.
Proposal
As a sort of short term emergency, I propose we introduce a fully manually curated public home feed.
Data Model
We add a new field
include_in_home_feed: boolean
to videos, make itfalse
by default.Query
We introduce a new query called
dumb_public_feed
which randomly selects up toSTATIC_HOME_FEED_LENGTH
videos so long asinclude_in_home_feed
is trueand returns them in random order, so expected to be different from query to query, even if data is same. I believe in PostgreSQL this query would look something like this
Operator API
We add two operator endpoints
include_in_home_feed
include_in_home_feed
for video with given ID.Operationally
Costs
The text was updated successfully, but these errors were encountered: