Rejected

:speech_balloon:

Don't automatically modify events after they're over

I notice that Arc will sometimes change the types or locations of previous events long after I’ve left, sometimes days later, when its model seems to think it knows something new.

This means that I can’t rely on any data in the past unless I’ve gone in and manually confirmed each event or even sample (even though the UI doesn’t show me what things have not been manually confirmed).

It would be better if Arc left things that I have previously seen alone, even when its learned model has changed. I look at timelines every few days, and correct errors, and I never want Arc to change something that I have reviewed.

I understand that for events that are in progress, more information might lead to a more accurate classification, and that’s desired. However, applying an updated model to old activities/places is never what I want.

1 vote

Tagged as Suggestion

Suggested 17 July 2019 by user Yinon

Moved into Rejected 17 July 2019

  • Sign in to comment and vote. Sign in by email
  • 17 July 2019 Yinon suggested this task

  • 17 July 2019 Matt Greenfield moved this task into Rejected

  • avatar

    Arc never makes changes to automatic place assignments after the initial assignment. However it will reclassify samples in trip items, if it has been a month since the samples were last classified and the samples don’t have a confirmed type.

    The reclassification should always be an improvement over the previous classification, because the ML models typically only ever improve in accuracy.

    If you have seen trips reclassified as a less accurate type than previously, then it would be good to have documentation of that, showing the before and after states, so that we can debug the problem and see why the new classification is less correct than the old. But that would require having both the before and after states available to assess, so that’s easier said than done 😉

    17 July 2019
  • avatar

    Can we add a reset feature for the machine learning if we go back and change older data substantially? All my moves imports are tagged as walking and my results are a bit buggy

    10 August 2019
  • avatar

    Tim, the ML models update for all of your changes. So when you go back and edit old data, those changes are incorporated into the models on next update (ie when your phone is next plugged in to power).

    There’s never a need to reset the models. Correcting any mistakes you find in the timeline will correct any mistakes in the models.

    11 August 2019