Machine learning-driven software architecture pattern recommendation: Bridging requirement specifications and design decisions

Abstract

This paper intends to present the Requirement-Aware Architecture Pattern Recommendation System that uses the Machine Learning(ML) as input with functional and non-functional requirement sets (FRs/NFRs) and automatically outputs a best-fit architecture pattern. The proposed approach is based on the use of deep natural language processing techniques (e.g., TF-IDF, word embedding) to identify meaningful features in requirement text. These features are then correlated with important quality attributes (scalability, security, performance, maintainability, availability etc.), and finally, supervised ML Classifiers (Random Forest, SVM, Decision Trees, NNs etc.) are employed to predict the most suitable architecture style. The system is trained and validated with-standard datasets, like PROMISE_exp, and also real-world Software Requirement Specification (SRS) documents. A demo-able, real-world prototype is also suggested, which would enable a user to put in his/her requirements and receive a real-time suggestion, justification, and a confidence score. This research will offer software architects, students and small-to-medium enterprises a tool to quickly, consistently and without bias select software architecture. The lack of work in the literature which discusses both NFR classification and architecture detection upon implementation of the system into source code will be addressed in the proposed approach; novelty comes from providing an end-to-end automated recommendation at the requirement level (pre-implementation stage).