Design patterns mining using neural sub-graph matching

Design Patterns detection in Object-Oriented software systems is essential for effectively supporting program comprehension and re-engineering tasks. It helps to recover, from source code, the developers' design decisions and trade-offs that could be not up-to-date or even not documented. Several approaches to mine design patterns from source code have been defined in the last twelve years and they are all based on the analysis of object-oriented systems components, their relationships, and behaviors to identify the roles played in the patterns. Both static and dynamic approaches need to perform matching between data captured from the system with the design patterns specification that encodes the structure and the behavior of the micro-architectural solution. The matching process, in principle, can be formulated as a sub-graph matching problem that is NP-complete. This problem has been addressed in the literature using heuristics designed to produce good solutions in an acceptable time, but the task is still expensive with a significant trade-off between accuracy and performance. This work proposes the adoption of a neural-based approach that exploits graph neural networks to perform detection using a more efficient sub-graph matching step outperforming existing heuristics proposed for this task. The pattern detection approach has been assessed on several open-source systems widely used to perform design pattern detection obtaining very good results for both detection performances and efficiency.