Abstract: Accurate software effort estimation (SEE) remains a challenging task due to missing values, structural inconsistencies, redundant attributes, and the heterogeneous nature of software engineering datasets. Although numerous machine learning models have been developed for effort estimation, comparatively less attention has been given to improving data quality and feature quality before model construction. As a result, estimation models are often trained on datasets containing incomplete, noisy, or redundant information, which can adversely affect prediction accuracy and model generalization. This study proposes an integrated software effort estimation framework that combines a Data Quality Enhancement Module (DQEM) with an Adaptive Multi-Criteria Redundancy-Aware Feature Optimization (AMRFO) framework. DQEM enhances dataset reliability through automated inspection, structural correction, adaptive missing-value reconstruction, uncertainty assessment, and diagnostic validation. The enhanced datasets are subsequently processed using AMRFO, which identifies informative, stable, and non-redundant feature subsets by jointly considering feature importance, feature stability, and feature redundancy. The proposed framework was evaluated on five benchmark software effort estimation datasets: Desharnais, China, ISBSG, COCOMO81, and Maxwell. Experimental evaluation using nine machine learning models and repeated cross-validation showed that AMRFO reduced the feature space by 50.00% to 72.22% across the evaluated datasets while maintaining or improving predictive performance in most cases. The largest improvement was observed for the COCOMO81 dataset, where the Mean Absolute Error (MAE) decreased from 454.66 to 320.09 after feature optimization. Overall, the results demonstrate that integrating data quality enhancement with multi-criteria feature optimization improves estimation reliability, reduces feature redundancy, and supports more accurate and generalizable software effort estimation. The proposed framework provides a practical and scalable solution for machine learning-based effort estimation using heterogeneous software engineering datasets.

Key words: software effort estimation, data quality enhancement, adaptive feature selection, feature stability, feature redundancy, machine learning, missing-value imputation, software engineering datasets