Assessing Simulated Software Graphs using Conditional Random Fields

In the field of software evolution, simulating the software development process is an important tool to understand the reasons why some projects fail, yet others prosper. For each simulation how- ever, there is a need to have an assessment of the simulation results. We use Conditional Random Fields, specifically a variant form based on the Ising model from theoretical physics, to assess software graph quality. Our CRF-based assessment model works on so called Software Graphs, where each node of that graph represents a software entity of the software project. The edges are determined by immediate dependencies between the pieces of software underlying the involved nodes. Because there is a lack of reference training data for our kind of evalua- tion, we engineered a special training paradigm that we call the Parsimo- nious Homogeneity Training. This training is not dependent on reference data. Instead of that it is designed to produce the following two effects. First, homogenizing the assessment of highly interconnected regions of the software graph, Second, leaving the assessment of these regions in relative independence from one another. The results presented demonstrate, that our assessment approach works.