Research On Graph Of Graphs Neural Network For Interaction Prediction Between Classes In Software System

Software,the end product of software engineering,is an extremely complex artefact that has undergone various rigorous design processes.As software evolves,the interactions between class files become more complex as the system grows in size.If the global structure of the software system is not properly adjusted to establish reasonable interactions between classes,development may deviate from the initial design and seriously affect the quality of the software.Therefore,it is a challenge for software maintenance to accurately predict the reasonable relationships between classes in a software system.In a software system,the relationships between classes can be modelled as a software network,where classes represent nodes and interactions between classes represent connected edges,and changes in structure during software evolution can be seen as the creation of new interactions and the removal of old ones.This study abstracts the software system into a graph of graphs structured software network based on the multi-granularity feature of software,and proposes a Graph of Graphs Neural Network-based Class Interaction Prediction(GoGCIP),the main work of this paper is as follows:(1)The internal graph representation in GoGCIP is investigated,and three types of semantic features are extracted from the micro and meso macro levels: character features,Token features and visual features,and seven sets of semantic features combined from the three types of semantics are used as the initial features of the external graph to verify the prediction effect of their combined approach.The experimental results show that the best prediction results can be achieved by using the fusion of all three types of semantic features,which has a significant advantage over other feature combination methods,with an average increase of 1.48% in the AUC value and 1.75% in the AP value.(2)The external graph neural network for GoGCIP was investigated,and three graph neural networks with different characteristics(GCN,Graph SAGE and GAT)were selected to be combined with seven sets of semantic features two by two to verify the most suitable external graph neural network for GoGCIP.The results show that using GAT as the external graph neural network achieves the best prediction results,with an average growth rate of0.97% for AUC values and 1.16% for AP values.(3)The prediction results on project datasets of different sizes based on different combinations of semantic features and external graph neural networks show that GoGCIP is suitable for predicting inter-class interactions in large and complex software systems.(4)For the two splicing methods of feature fusion,the use of merged splicing was verified to be more effective for interaction prediction on GoGCIP.