Towards Generating Structurally Realistic Models by Generative Adversarial Networks

Several activities in model-driven engineering (MDE), like model transformation testing, would require the availability of big sets of realistic models. However, the community has failed so far in producing large model repositories, and the lack of freely available industrial models has been raised as one of the most important problems in MDE. Consequently, MDE researchers have developed various tools and methods to generate models using different approaches, such as graph grammar, partitioning, and random generation. However, these tools rarely focus on producing new models, considering their realism. Contribution. In this work, we utilize generative deep learning, in particular, Generative Adversarial Networks (GANs), to present an approach for generating new structurally realistic models. Built atop the Eclipse Modeling Framework, the proposed tool can produce new artificial models from a metamodel and one big instance model as inputs. Graph-based metrics have been used to evaluate the approach. The preliminary statistical results illustrate that using GANs can be promising for creating new realistic models.