Page MenuHomePhabricator

Bias detection
Open, HighPublic

Description

User interactions are used across our systems to measure whether a post could be relevant to the community. The issue we're encountering is that users will often try to break these mechanisms and attempt to leverage them to their advantage.

Ping is a central component to ensuring that the system stays healthy and that content that is actually relevant is properly featured. There's two mechanisms that I would like to introduce here:

  • Bias detection. To prevent bad actors from creating "noise" that is intended to distract the system. For example, sock puppet accounts and like-bombing are a common theme.
  • Personalized recommendations I would like to drive the system towards personalization, so that a user sees featured content that is relevant within their circles whenever possible. The issue here is that this can be computationally expensive.

These should help discourage users who wish to "poison the well" by spamming likes and comments on content that is not relevant to other users on the platform.

Related Objects

Event Timeline

cesar triaged this task as High priority.Mar 27 2020, 4:31 PM
cesar created this task.
cesar created this object with visibility "Public (No Login Required)".
cesar added a comment.Mar 27 2020, 4:38 PM

I'd love to see a mechanism along the lines of defer being used for calculating the new scores. This could provide something like:

  • When a user logs in, check whether recommendations are stale. If they are, create a task to generate new ones as soon as possible.
  • Whenever the user navigates the page, and their recommendations start to become stale, create a task to regenerate the stale ones.
    • Also, create tasks to recalculate stale-ness every few hours.
    • Create a task to recalculate recommendations in a week and send them via notification (this task should only be executed if the user was not seen since)
cesar added a comment.Apr 6 2020, 10:58 AM

I created T120 to document the efforts on personalized recommendations.

I did not add them as a subtask, because, while it's probably instrumental to spam mitigation to have a system that is by design, immune to the effects of spam, it is not a dependency for this project to advance.