thesis - Faculty of Information and Communication Technologies ...

Chapter 6. Change Dynamics
new classes with that of all classes, we can see (Figure 6.7) that this
proportion is consistently lower than the norm. New classes, therefore,
tend to start out with relatively lower popularity and some of these gain
dependents over time. The observation that in general new classes are
less popular than existing classes is also supported by our finding from
the Gini coefficient analysis in the previous chapter. We stated that the
In-Degree Count gini coefficient in general increases as systems mature
which also shows that new classes do not have the same distribution
profile in terms of popularity as the existing classes — if they did, the
Gini Coefficient would not change.
6.2.7 Structural Complexity of Modified Classes
In the previous section we have shown that the popularity of a class
makes it more change-prone. But, does the internal structural complexity
of a class increase the likelihood that a class will be modified?
In this section, we address this question by investigating the relationship
between the Number of Branches and change.
When a class is modified, there is a certain probability that the branching
statements are altered. To appreciate if the number of branching
instructions has an impact on the probability of change, we constructed
two logistic regression models. In the first model, Number of Branches
was the explanatory (independent) variable. For the second model, we
used the size normalized Number of Branches as the explanatory variable.
The size normalized Number of Branches is a ratio calculated as
the number of branches per bytecode instructions in a class. The size
normalization was applied to identify if size causes a significant distortion
in the probability of modification. That is, the class is being modified
because it is bigger rather than because it has more structural
complexity. In both models the response (dependent) variable was if a
class was modified at least once in its release history (0 was used to
represent classes that were not modified, 1 indicated class was modified
at least once in its evolution history). The logistic regression models
were constructed by using the set of all classes across all versions of all
software systems in our input data set.
162