13. AD Data Quality Queen

Why is data quality especially important for machine learning?

ML performance depends strongly on the quality of the training data. Data that does not satisfy the requirements of the intended task can produce inaccurate and failure-prone models, so data quality must be defined, measured, and managed throughout the data life cycle.

Name and describe important ISO/IEC 5259-2 data-quality characteristics.

Accuracy means data values are correct; completeness means expected attributes and instances are present; consistency means data is coherent and free of contradictions; credibility means data is considered believable in its context of use; and currentness means the data has an appropriate age for its intended use.

What are the three main principles of the QUEEN method?

First, reduce task complexity through preprocessing and use ML mainly where it provides a benefit. Second, use quality indicators to measure global and local mapping complexity and guide this simplification. Third, apply ML methods that permit glass-box evaluation so that systematic errors in the learned mapping can be detected.

What is the basic development idea behind QUEEN?

QUEEN tries to simplify a nonlinear learning task step by step so that the structure in the input space increasingly corresponds to the structure in the output space. It compares pairs of examples and uses quality indicators to evaluate whether transformations and encodings reduce the complexity of the mapping

What is an Equivalence Class Set (ECS) and how is equivalence determined?

ECS groups pairs of data points according to whether they are equivalent or unequivalent in input and output space. Equivalence is task-specific and is determined using suitable distance metrics and thresholds for the input and output spaces.

What do the four ECS classes EE, EU, UE and UU mean?

The first letter refers to the input space, the second to the output space: EE = equivalent input and equivalent output; EU = equivalent input but unequivalent output; UE = unequivalent input but equivalent output; UU = unequivalent in both spaces.

Why are ECS-EU pairs particularly relevant for data-quality analysis?

EU means two examples are very similar in the input space but have substantially different outputs. Such pairs can indicate outliers, inconsistent labels, or problematic decision regions, making them especially useful for detecting data-quality problems.