[ML] Change point detection and prediction #92
Merged
Conversation
This file contains 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
…s of the name prior
…tection and modelling (#9) This implements 1) a naive Bayes classifier, using our distribution models, which will be used for modelling the probability of a change, and 2) a change detector framework, currently supporting detecting level shifts and time shifts, which works by comparing BIC of the various possible hypotheses against one another and a null hypothesis that there is no change.
…ns of c library functions in maths CTools
… model (#11) This wires in change detection and starts some unit testing of CTimeSeriesModel with change points. There is some more work to be done to avoid using level shifts to try and fit other types of change points, such as scaling.
This implements detection of linear scaling events. It also finishes up the unit testing of change detection and fixes some issues these turned up: specifically, 1) the behaviour when a change is detected but the trend model has no components, 2) the handling of time shifts in the trend model and 3) the handling of data types in the trend component change model. Finally, we are now more careful with the weights we apply to samples added to both the standard and change models. This has meant I've been able to revert scaling the changes, since the trend is less influenced by values during the change detection period if we're likely to detect a change.
…forecast-enhancements-part-2
…forecast-enhancements-part-2
…forecast-enhancements-part-2
…forecast-enhancements-part-2
…stic/ml-cpp into feature/forecast-enhancements-part-2
tveasey
added a commit
that referenced
this pull request
May 22, 2018
droberts195
changed the title
This was referenced Dec 18, 2018
Merged
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.
This implements a mechanism to detect specific types of change points in a time series; specifically, 1) step change, 2) linear scaling and 3) time shift (e.g. for daylight saving). These are then applied to the time series model in one shot (rather than gradually relearning its parameters). In addition, it introduces a new probabilistic model to predict when step changes might occur in the future and uses this to roll out possible future paths for the time series including steps.
The fact that we can now detect specific types of change points has also meant that we can increase the model robustness w.r.t. outliers, since behaviour at significant change points doesn't as strongly rely on being able to adjust the model parameters with unusual values.
From an anomaly detection perspective this means we can adjust the model with significantly less growth in prediction uncertainty when the time series undergoes one of the modelled class of changes. This results in fewer blind spots in detection for such signals. From a forecasting perspective, signals with predictable discontinuities, such a garbage collection, are much more accurately predicted.
This PR commits the long running task on change point detection and modelling to master. Note that these changes have already been individually reviewed.