• 0ops@lemm.ee
    link
    fedilink
    English
    arrow-up
    3
    ·
    edit-2
    5 months ago

    Right, there’s typically separate “training” and “validation” sets for a model to train, validate, and iterate on, and then a totally separate “test” dataset that measures how effective the model is on similar data that it wasn’t trained on.

    If the model gets good results on the validation dataset but less good on the test dataset, that typically means that it’s “over fit”. Essentially the model started memorizing frivolous details specific to the validation set that while they do improve evaluation results on that specific dataset, they do nothing or even hurt the results for the testing and other datasets that weren’t a part of training. Basically, the model failed to abstract what it’s supposed to detect, only managing good results in validation through brute memorization.

    I’m not sure if that’s quite what’s happening in maven’s description though. If it’s real my initial thoughts are an unrepresentative dataset + failing to reach high accuracy to begin with. I buy that there’s a correlation between machine specs and positive cases, but I’m sure it’s not a perfect correlation. Like maven said, old areas get new machines sometimes. If the models accuracy was never high to begin with, that correlation may just be the models best guess. Even though I’m sure that it would always take machine specs into account as long as they’re part of the dataset, if actual symptoms correlate more strongly to positive diagnoses than machine specs do, then I’d expect the model to evaluate primarily on symptoms, and thus be more accurate. Sorry this got longer than I wanted

    • Tja@programming.dev
      link
      fedilink
      English
      arrow-up
      1
      arrow-down
      1
      ·
      5 months ago

      It’s no problem to have a longer description if you want to get nuance. I think that’s a good description and fair assumptions. Reality is rarely as black and white as reddit/lemmy wants it to be.