Data doping

As explained in Outlier Detection in Python

Data doping is the deliberate injection or alteration of observations in a dataset to test a detector under controlled conditions.

Data doping creates test observations by copying real records and deliberately changing selected values. The aim is to simulate outliers with known modifications, then check whether a detector scores them above unmodified records. Unlike generating entirely new rows, doping preserves much of each record’s original context. Its usefulness depends on whether the changes resemble the anomalies worth finding. section 8.5.1 section 11.3

Changing a typical value into an unusual combination

Copy a tabular dataset and select one cell per row at random. Replace a value above its feature’s median with a sampled value below the median, or vice versa. The replacement can still be an ordinary value on its own, while becoming unusual alongside the row’s other features. Fit detectors on real data and compare their scores for real and doped rows: higher doped scores indicate sensitivity to this particular kind of disruption, not to every possible outlier. section 8.5.1

Figure 8.10 Scores for real versus doped records when tested on IF, LOF, and ECOD. For this form of doping, LOF performs very well and IF reasonably well. As no extreme values were added in this example, ECOD makes no distinction between the real and doped data.

The score comparison shows why the modification matters: LOF separates these doped records well, IF does so reasonably well, and ECOD does not distinguish them when no extreme values are introduced. section 8.5.1

Design several controlled tests rather than treating one doping rule as a universal benchmark:

• Vary how many features change, how far values move, and whether replacements may exceed observed numeric ranges or categorical values. section 11.3
• Use a consistent numeric shift when a known change magnitude is useful; create test sets at different difficulty levels. section 11.3
• Include relevant patterns such as extreme values, moderate changes across multiple features, and rare combinations; exclude patterns outside the detection goal. section 11.3

Interpret results against the chosen modifications. A changed cell may disturb several feature associations, so its effect is not fully specified by the number or size of edits. Inspect flagged records as well as scores: visually ordinary doped rows need not all be flagged, while conspicuously unusual combinations should be. Doped rows can also be mixed in small numbers with current data to test detection there, provided they do not substantially change the training distribution. Ground-truth labels make metrics such as AUROC possible, but hand-labeling and project-specific doping remain more relevant than general detector rankings. Avoid evaluating a system solely with the same synthetic-data method it uses to identify outliers. section 8.5.1 section 8.8 section 11.3 section 11.6

How Outlier Detection in Python covers it

How deep it goes

Works through a median-crossing modification and detector comparison, then specifies controllable changes and cautions about disrupted feature associations.

Examples

Compares IF, LOF, and ECOD on median-crossing doped records and illustrates unusual Length–Diameter combinations.

What it assumes

Assumes familiarity with tabular feature values and detector scores; defines doping through a concrete row-modification procedure.

What you can do after

Practice/perform it by modifying real rows at controlled difficulty levels and comparing detector scores with scores for unmodified rows.

How it connects

Connects synthetic data, outlier detection, detector ensembles, ground-truth labels, and AUROC.

liveBookmanning.com

Read Outlier Detection in Python on liveBook Get the book at manning.com

Related concepts

• outlier detectionuses · Outlier detection is the process of identifying observations that substantially deviate from an expected or prevailing pattern in data.
• random forestused by · A random forest is an ensemble model that combines predictions from multiple decision trees built with randomized data samples, feature selections, or both.
• synthetic dataproduces · Synthetic data is data generated artificially rather than collected directly from the target process, often to approximate its properties or meet specified requirements.
• accuracymeasures · Accuracy is the degree to which an output, prediction, or measurement agrees with a correct or reference outcome under a specified criterion.
• ensembleuses · An ensemble is a technique that combines or considers outputs from multiple models or component methods collectively to produce or assess a joint result.
• area under the curveused by · The area under a curve is the integral of its values over a specified domain and, in classification evaluation, commonly summarizes performance across decision thresholds.

This explainer was generated by Manning's knowledge graph from the text of Outlier Detection in Python. The section links point to the parts of the book it draws on.