Drift Simulation Overview

Pucktrick supports simulation of three structurally distinct forms of dataset drift, modelled as temporal corruption policies applied to dataset segments. Drift is useful for evaluating machine learning models under non-stationary data conditions.

Function Signature

from pucktrick.drift import drift

error, df_modified = drift(df, strategy)

Unlike other modules, the error return value is a dictionary:

{
    "errore": "yes",                     # "yes" if any modification occurred, "no" otherwise
    "change_points": [100, 200, 300],    # row indices delimiting chunk boundaries
    "chunks": {
        "0": {"start": 0,   "end": 100, "drift_applied": False},
        "1": {"start": 100, "end": 200, "drift_applied": False},
        "2": {"start": 200, "end": 300, "drift_applied": True},
        "3": {"start": 300, "end": 400, "drift_applied": True},
    }
}

Strategy Format

The drift strategy follows the same structure as all other Pucktrick modules. Chunk configurations are placed inside perturbate_data:

strategy = {
    "affected_features": ["f1", "f2"],
    "selection_criteria": "all",
    "percentage": 0.35,          # fraction of rows affected per chunk
    "mode": "new",
    "perturbate_data": {
        "sampling": "random",
        "target_col": "target",  # omit for auto-detection
        "chunks": {
            "0": None,           # baseline segment, no drift
            "1": None,
            "2": { ... },        # drift configuration for segment 2
            "3": { ... },
        }
    }
}

Note: sampling and distribution parameters inside perturbate_data are not used by the drift module — row selection within each chunk is always random and controlled exclusively by percentage.

Drift Types Summary

Drift type	Distribution affected	Module	drift_type value
Data drift (covariate noise)	$P(X)$ changes	covariate_noise_drift	"covariate_noise"
Data drift (offset)	$P(X)$ changes	covariate_offset_drift	"covariate_offset"
Concept drift (target offset)	$P(Y\|X)$ changes	offset_drift	"concept"
Concept drift (feature rotation)	$P(Y\|X)$ changes	concept_drift	"concept_rotation"
Label drift (prior shift)	$P(Y)$ changes	prior_multinomial_drift	"prior_multinomial"
Target scaling	$P(Y)$ changes	target_scaling_drift	"target_scaling"
Generic (all types)	configurable	drift_generic	any