hesela.devSMOTE
SMOTE is an oversampling method that creates synthetic minority-class feature vectors by interpolating between a minority example and selected minority-class nearest neighbors.
For numeric features, the original construction uses x_new = x + u * (x_neighbor - x), with u drawn between 0 and 1; it creates feature-space samples, not new observed drive failures. [1] Fit resampling inside each training fold, leaving validation and test observations unresampled; resampling before splitting can leak information and distort evaluation prevalence. [2] A sampler pipeline does not itself establish an appropriate chronological or device-group split for storage telemetry. [3]
Relations
- related
- class-prevalence, precision, recall, smart
Limitations
- Interpolation assumes a meaningful feature representation and neighborhood; a synthetic SMART vector is not evidence that a physical drive followed that trajectory. [1]
- Oversampling is a training choice, not a validation protocol or a guarantee of improved deployed precision. [2, 3]