Recommender system for model driven software development

Abstract

Models are key artifacts in model driven software engineering, similar to source code in traditional software engineering. Integrated development environments help users while writing source code, e.g. with typed auto completions, quick fixes, or automatic refactorings. Similar integrated features are rare for modeling IDEs. The above source code IDE features can be seen as a recommender system. A recommender system for model driven software engineering can combine data from different sources in order to infer a list of relevant and actionable model changes in real time. These recommendations can speed up working on models by automating repetitive tasks and preventing errors when the changes are atypical for the changed models. Recommendations can be based on common model transformations that are taken from the literature or learned from models in version control systems. Further information can be taken from instance- to meta-model relationships, modeling related artifacts (e.g. correctness constraints), and versions histories of models under version control. We created a prototype recommender that analyses the change history of a single model. We computed its accuracy via cross-validation and found that it was between 0.43 and 0.82 for models from an open source project. In order to have a bigger data set for the evaluation and the learning of model transformation, we also mined repositories from Eclipse projects for Ecore meta models and their versions. We found 4374 meta models with 17249 versions. 244 of these meta models were changed at least ten times and are candidates for learning common model transformations. We plan to evaluate our recommender system in two ways: (1) In off-line evaluations with data sets of models from the literature, created by us, or taken from industry partners. (2) In on-line user studies with participants from academia and industry, performed as case studies and controlled experiments.