Human-Centric Machine Learning Frameworks for Scalable Software Quality Prediction

The human-centred perspective of software quality embraces the construction of data-driven software quality prediction models, which inherently support quality governance decisions over the life cycle. While the introduced models are directly linked to software engineering quality dimensions, they lack a well-established logical relationship with stakeholder-centric software quality attributes. A conceptual framework that explicitly links quality-attribute drivers to human-centric software quality dimensional data suitable for software quality prediction has yet to be defined. The proposed framework addresses this gap and provides a structured approach for developing data sources and human-centric software quality prediction models. It supports the construction of prediction models in line with stakeholder requirements, enables the selection of quality indicators focused on business and operational objectives, and strengthens the connection between quality models and quality dimensions aimed at regulatory compliance, stakeholder trust, and attractiveness. This extended version contributes twelve numbered equations that formalise stakeholder weighting, trust, calibration, and scalability; four algorithms covering elicitation, trust-gated model routing, explanation-driven trust calibration, and governance-staged rollout; and an empirical section grounded in 2025 practitioner evidence on trust in AI-generated code.In the last two decades, machine learning and artificial intelligence have enabled new research and practical applications in various domains. However, the construction of data-driven models for predicting the quality of software products that generate scalable outcomes aligned with the requirements of those who make quality decisions throughout the life cycle remains an open issue. The necessity to develop human-centric software, recognise quality from a human-centric perspective, and comply with regulations and corporate governance principles have become important drivers for research and practice in software engineering. Accordingly, the demand for open and interpretable machine learning models is steadily growing. Yet the deep-learning nature of built models limits their direct application and restricts their operationalisation. These aspects must be considered not only in model construction or training but throughout the complete cycle of model conception, construction, and use.