<p>Software change proneness prediction in Android applications is crucial in enhancing software maintenance and evolution processes. Inadequate consideration of data imbalance may compromise the reliability and robustness of change proneness prediction models. The study proposes a novel approach that leverages advanced techniques to address data imbalance while improving predictive accuracy. The methodology integrates state-of-the-art techniques to address key challenges. SMOTE (Synthetic Minority Over-sampling Technique) is employed for handling class imbalance, Self-Attention and RoBERTa embeddings are used to extract rich contextual information from text inputs, and Recursive Feature Elimination (RFE) is applied for feature selection to enhance model interpretability and performance. Moreover, we employ XLNet coupled with Butterfly optimization to reduce model weight while maintaining high predictive accuracy effectively. Our hybrid approach offers a robust framework for software change proneness prediction in Android applications by combining advanced language representation learning with optimization techniques. The efficiency of the presented technique is analyzed based on various metrics and research work compared with different state-of-the-art works.</p>