Permalink
Branch: master
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
42 lines (29 sloc) 2.02 KB

Score Normalization and Stable Model Updates

What is score normalization for?

Score normalization of our models is intended to do two things:

  • Provide a way to interpret the scores of a model as a probability (e.g. how likely is it that the comment will be perceived as toxic).
  • Enable models to be improved without any action needed by most clients, Without normalization, clients using the model would have to change the thresholds they use each time our models are retrained with additional examples.

When did score-normalizing start happening?

How is score normalization done?

Our normalization process applies a version of probability calibration by isotonic regression to make the scores approximate probabilities. For example, so that a toxicity score of 0.80 can be interpreted as a belief that 80% of people would consider the comment to be toxic.

However, as we introduce additional data, the basic distribution of how many toxic comments there are can change. Because models are far from perfect, this would result in a shift in the calibration scores. To control for this, we perform callibration on a randomly selected 50/50 class split. of course, this control is not perfect: if the newly introduced examples are differently difficult to previous ones, some change in scores is still likely. However, in practice, we think this will be reasonably stable, and clients will be able to have the same 'sense' of what a threshold means.

We are always looking for better ways to provide stability, so if you know of better ways to do this, please do suggest it on github.