HumVDetClas: A context-aware heterogeneous ensemble for detecting and classifying human value violations in app reviews

Mobile applications are integral to daily life, making end-user considerations and human values crucial in software development. Research shows that violating human values in mobile apps can lead to data breaches, legal issues, and user dissatisfaction. We propose a contextual two-tieredhuman values detection and classification (HumVDetClas) approach analyzing end-user reviews of low-rated Amazon App Store applications. We collected 78,213 reviews from 59 low-ranked mobile applications to explore human value violations. A random sample of 400 reviews was analyzed using Schwartz’s theory to identify human value violations. Using this theory and content analysis, we annotated 18,256 user comments to develop a dataset of human value violations. We fine-tuned Deep Learning (DL) and transformer classifiers to detect violations and classify them into ten types aligned with Schwartz’s theory. The HumVDetClas approach shows fine-tuned BERT and DistilBERT achieve 94.59% and 94.19% binary test accuracy, and 77.16% and 76.34% multi-class F1-score, outperforming DL baselines. Next, we introduce a heterogeneous ensemble combining CNN, BiGRU, DistilBERT, and BERT via soft voting fusion with automatic class remapping that attains 79.34% multi-class F1-score, an improvement over BERT by 2.2%. HumVDetClas surpasses keyword-based approaches through deep contextual understanding and can be integrated into software evolution pipelines to address ethical concerns and enhance user trust.