Type Inference in Flexible Model-Driven Engineering

Abstract

Model-driven Engineering (MDE) is an approach to software development that promises increased productivity and product quality. Domain models that conform to metamodels, both of which are the core artefacts in MDE approaches, are manipulated to perform different development processes using specific MDE tools.However, domain experts, who have detailed domain knowledge, typically lack the technical expertise to transfer this knowledge using MDE tools. Flexible or bottom-up Model-driven Engineering is an emerging approach to domain and systems modelling that tackles this challenge by promoting the use of simple drawing tools to increase the involvement of domain experts in MDE processes. In this approach, no metamodel is created upfront but instead the process starts with the definition of example models that will be used to infer a draft metamodel. When complete knowledge of the domain is acquired, a final metamodel is devised and a transition to traditional MDE approaches is possible. However, the lack of a metamodel that encodes the semantics of conforming models and of tools that impose these semantics bears some drawbacks, among others that of having models with nodes that are unintentionally left untyped. In this thesis we propose the use of approaches that use algorithms from three different research areas, that of classification algorithms, constraint programming and graph similarity to help with the type inference of such untyped nodes. We perform an evaluation of the proposed approaches in a number of randomly generated example models from 10 different domains with results suggesting that the approaches could be used for type inference both in an automatic or a semi-automatic style. 3For my parents Despoina and Michalis

Type inference in flexible model-driven engineering using classification algorithms

“…6 To evaluate our approach, we applied it to a number of randomly generated models, instances of publicly available metamodels that were collected as part of the work presented in [37]. The 10 metamodels selected are the same used in the evaluation of our previous work [43]. For each of these metamodels, we produced 10 random instances using the Crepe model generator tool [36] (step 1 in Fig.…”

“…In this approach, we use a set of five features, presented in Table 2 which are the same as those used in our previous work [43]. These features were selected because they arguably measure structural and semantic characteristics of the models.…”

“…our feature signature [3,3,1, Many classification algorithms exist, some of the most established being decision trees, random forests, support vector machines and neural networks [15]. For our previous work [43], we chose to use decision trees due to the interpretable output representing the hypothesis learnt. In practice, other classification algorithms can often have higher accuracy, but will produce a hypothesis in a form that is not human readable.…”

“…In addition, elements can be created without having a type assigned to them, which can be calculated when it suits the domain expert, not the modelling tool. This paper builds on top of the work presented in [43]. In [43], we proposed the use of a classification algorithm, specifically classification and regression trees (CART) [4], to calculate matches between typed and untyped elements, based on five characteristics of each of them.…”

“…This paper builds on top of the work presented in [43]. In [43], we proposed the use of a classification algorithm, specifically classification and regression trees (CART) [4], to calculate matches between typed and untyped elements, based on five characteristics of each of them. We demonstrated the approach using a flexible modelling technique based on GraphML, called Muddles [18].…”