Using A Transformer To Evaluate The Impact Of Architectural Smells On Software Performance

Architectural smells—design flaws in software—can significantly degrade system performanceby increasing execution time, memory consumption, and reducing overall efficiency. While exist-ing tools use static analysis to detect these smells, their impact on software maintainability andrefactoring strategies remains largely unexplored. Traditional approaches, such as rule-based orstatistical models, struggle to recommend effective refactoring actions due to the complexity ofarchitectural dependencies and design decisions.This paper introduces a novel transformer-based model designed to predict the most suitablerefactoring actions for addressing architectural smells. By leveraging sequence-to-sequence learning,the model captures dependencies between architectural patterns and optimal refactoring strategies,enabling more precise and context-aware recommendations. Integrating deep learning into archi-tectural evaluation overcomes the limitations of traditional methods and provides deeper insightsinto software maintainability and evolution.The model is evaluated on a dataset of open-source projects, analyzing common architecturalsmells such as god classes, cyclic dependencies, and hub-like dependencies. Results demonstratethat it outperforms conventional models in suggesting refactoring that improve key software qual-ity attributes, such as modularity, maintainability, and performance, with higher accuracy androbustness.This study contributes to automated architectural assessment by bridging the gap between soft-ware design flaws and refactoring strategies. The findings highlight the potential of machine learningin improving software architecture, offering developers a powerful tool for proactive code improve-ment and long-term software sustainability.