.png){kind=icon}
.png){kind=icon}
Return
Link Prediction for Completing Graphical Software Models Using Neural Networks
.png){kind=icon}
OA
AI
DOI：10.1109/ACCESS.2023.3323591.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deficiencies and inconsistencies introduced during the modeling of software systems may result in high costs and negatively impact the quality of all developments performed using these models. Therefore, developing more accurate models will aid software architects in developing software systems that match and exceed expectations. This paper proposes a graph neural network (GNN) method for predicting missing connections, or links, in graphical models, which are widely employed in modeling software systems. The proposed method utilizes graphs as allegedly incomplete, primitive graphical models of the system under consideration (SUC) as input and proposes links between its elements through the following steps: (i) transform the models into graph-structured data and extract features from the nodes, (ii) train the GNN model, and (iii) evaluate the performance of the trained model. Two GNN models based on SEAL and DeepLinker are evaluated using three performance metrics, namely cross-entropy loss, area under curve, and accuracy. Event sequence graphs (ESGs) are used as an example of applying the approach to an event-based behavioral modeling technique. Examining the results of experiments conducted on various datasets and variations of GNN reveals that missing connections between events in an ESG can be predicted even with relatively small datasets generated from ESG models.
Keywords:
Software engineering
Predictive models
Graph neural networks
Graphical user interfaces
Graphical models
Data models
Behavioral sciences
Event detection
Couplings
Event-based modeling
graph neural networks
link prediction