Information Technology Reference
In-Depth Information
imperativewhen developingCMAmethods and is hence rarely evident fromscientific
publications. For that reason it is not further investigated in the following.
3.3.1 Validation
Validation investigates if and how well a new method does what it is supposed to do.
Often this involves a comparison of method outcomes with respective observations or
manual measurements made by experts. Rykiel (
1996
) lists a wide range of different
validation tests. In the context of this chapter five shall be portrayed in more detail.
3.3.1.1 Face Validity
For assessing the face validity of a method knowledgeable people are asked if the
method and its behavior is reasonable, if input-output relationships appear reason-
able. For instance, discussions with farming experts when working on Laube et al.
(
P12
.
2011a
) revealed that the tool-driven definition of the movement pattern
flock
was not optimal. The pattern definition used required the individuals to move within
a circular disc of a given radius, whereas the observed movement rather showed
flocks as chains of piecewise connected pairs. Here, the face validity test revealed a
limited suitability of the chosen formalization underlying the proposed method.
3.3.1.2 Visualization Techniques
Another validation strategy is offered by visualization, exploratory analysis, or visual
analytics, where the data mining process is combined with a human analyst. Here the
user directly inspects the method outcomes by the use of visual displays and thereby
validates the plausibility of method outcomes. Examples for visual validation can be
found in Sect.
3.2.4
.
3.3.1.3 Internal Validity
For assessing the internal validity of a method test data sets can be used for investi-
gating if the method produces and reproduces a consistent output. The error analysis
carried out for decentralized flock mining in Laube et al. (
P6
.
2008b
) and Laube
et al. (
P12
.
2011a
) shall serve here as an example for testing the internal validity.
Given a data set with a known spatio-temporal occurrence of target patterns, the num-
ber of actually present patterns is compared to the number of patterns found by the
movement mining algorithm.
Error of omission
(eoo, “missed patterns”) accounts
for existing patterns not found, while
error of commission
(eoc, “false positives”)
specifies the wrongly detected patterns when no pattern actually exists.