This PR enhances ARFAdapter and fixes few details in DriftModelAdapterBase, mostly:
ARFAdapter now tracks:
mean metrics for all base tree attributes (river.tree.HoeffdingTree), e.g. tree height, number of leafs, etc.
mean, max and min values for selected metric used for weighted voting
ADWINAdapter parameters per tree (so there are 2 x n_models x adwin_params charts created for drift and warning detection
DriftModelAdapterBase now distinguishes between models with 2 detectors (warning & drift) and combined (like in CBCE - DDM)
With current implementation, there might be problems with "non-standard" params selection for ARFClassifier:
background trees are disabled if warning_drift_detector is disabled, there might be problems if adapter is used when NoDrift() is used
weighted-voting metrics are tracked even if voting is disabled
In future this might be bullet-proofed, but currently I don't think we have time for that. If we will have any problems with the future experiments and certain parameters, I believe we can quickfix it.
This PR enhances
ARFAdapterand fixes few details inDriftModelAdapterBase, mostly:meanmetrics for all base tree attributes (river.tree.HoeffdingTree), e.g. tree height, number of leafs, etc.mean,maxandminvalues for selectedmetricused for weighted votingADWINAdapterparameters per tree (so there are 2 x n_models x adwin_params charts created for drift and warning detectionDriftModelAdapterBasenow distinguishes between models with 2 detectors (warning & drift) and combined (like in CBCE - DDM)With current implementation, there might be problems with "non-standard" params selection for ARFClassifier:
backgroundtrees are disabled if warning_drift_detector is disabled, there might be problems if adapter is used whenNoDrift()is usedweighted-votingmetrics are tracked even if voting is disabledIn future this might be bullet-proofed, but currently I don't think we have time for that. If we will have any problems with the future experiments and certain parameters, I believe we can quickfix it.