[ML] Address the root cause for "actual equals typical equals zero" anomalies #2270
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
droberts195
approved these changes
May 17, 2022
There was a problem hiding this comment.
LGTM
The overall approach looks good. I don't know the code well enough to say whether the truncation has been applied everywhere where it needs to be. Probably the biggest chance of a mistake in this PR is a line that hasn't been changed rather than one that has, but at least that would be no worse than now.
…lly calling the wrong implementation
tveasey
added a commit
that referenced
this pull request
Jul 27, 2023
…nomalies (#2270) We've had multiple issues reported in the past where we report high anomaly scores when the actual is the same as the typical and both are equal to zero. This has been exacerbated in the past because we've also had some instabilities in the modelling when we stop receiving data for a partition, which was an edge case when this could occur. In all problematic cases, the underlying reason is the if we know the data are non-negative we truncate the prediction when we report the typical value. However, up until now we haven't truncated when computing probabilities. This changes switches to always truncating our prediction. (This is a halfway house to a more rigorous approach which would condition the predicted distribution to be non-negative.) This should be good enough to ensure we never create non- zero anomaly scores for cases where actual equals typical equals zero except when the result is a multi-bucket anomaly.
tveasey
added a commit
that referenced
this pull request
Jul 27, 2023
…nomalies (#2270) We've had multiple issues reported in the past where we report high anomaly scores when the actual is the same as the typical and both are equal to zero. This has been exacerbated in the past because we've also had some instabilities in the modelling when we stop receiving data for a partition, which was an edge case when this could occur. In all problematic cases, the underlying reason is the if we know the data are non-negative we truncate the prediction when we report the typical value. However, up until now we haven't truncated when computing probabilities. This changes switches to always truncating our prediction. (This is a halfway house to a more rigorous approach which would condition the predicted distribution to be non-negative.) This should be good enough to ensure we never create non- zero anomaly scores for cases where actual equals typical equals zero except when the result is a multi-bucket anomaly.
tveasey
added a commit
that referenced
this pull request
Jul 27, 2023
…nomalies (#2270) We've had multiple issues reported in the past where we report high anomaly scores when the actual is the same as the typical and both are equal to zero. This has been exacerbated in the past because we've also had some instabilities in the modelling when we stop receiving data for a partition, which was an edge case when this could occur. In all problematic cases, the underlying reason is the if we know the data are non-negative we truncate the prediction when we report the typical value. However, up until now we haven't truncated when computing probabilities. This changes switches to always truncating our prediction. (This is a halfway house to a more rigorous approach which would condition the predicted distribution to be non-negative.) This should be good enough to ensure we never create non- zero anomaly scores for cases where actual equals typical equals zero except when the result is a multi-bucket anomaly.
tveasey
added a commit
that referenced
this pull request
Jul 27, 2023
…nomalies (#2270) We've had multiple issues reported in the past where we report high anomaly scores when the actual is the same as the typical and both are equal to zero. This has been exacerbated in the past because we've also had some instabilities in the modelling when we stop receiving data for a partition, which was an edge case when this could occur. In all problematic cases, the underlying reason is the if we know the data are non-negative we truncate the prediction when we report the typical value. However, up until now we haven't truncated when computing probabilities. This changes switches to always truncating our prediction. (This is a halfway house to a more rigorous approach which would condition the predicted distribution to be non-negative.) This should be good enough to ensure we never create non- zero anomaly scores for cases where actual equals typical equals zero except when the result is a multi-bucket anomaly.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
We've had multiple issues reported in the past where we report high anomaly scores when the actual is the same as the typical and both are equal to zero. This has been exacerbated in the past because we've also had some instabilities in the modelling when we stop receiving data for a partition, which was an edge case when this could occur.
In all problematic cases, the underlying reason is that if we know the data are non-negative we truncate the prediction when we report the typical value. However, up until now we haven't truncated when computing probabilities.
This change switches to always truncating our prediction. (This is a halfway house to a more rigorous approach which would condition that the predicted distribution is non-negative.) This should be good enough to ensure we never create non-zero anomaly scores for cases where actual equals typical equals zero except when the result is a multi-bucket anomaly.