Machine Learning based Effort Estimation Models for Software Development Projects related Datasets with diverse features

This research evaluates the performance of machine learning models for the estimation of software effort using diverse datasets with distinct feature sets and varying training-test splits. To address the risks of under or overestimation in project management, the study employs robust data preprocessing, feature engineering, and selection techniques. Through extensive experimentation, the study analyzes model accuracy using multiple metrics such as Mean Absolute Error, R-squared, precision, recall, F1-score, and Receiver Operating Characteristic Area Under the Curve. Preprocessing techniques, including feature scaling and outlier removal, were applied to reduce overfitting and improve generalization. Performance was assessed using various indicators, with results showing that the stacked ensemble model consistently outperformed other models like Linear Regression, Support Vector Regression, and Gradient Boosting in generating accurate predictions across datasets. The analysis revealed that no single data split yielded optimal performance for all datasets; instead, different datasets performed best under customized splits. Larger datasets performed well with splits like 70-30, while smaller ones exhibited overfitting even after preprocessing. Models like Random Forest, XGBoost, and Gradient Boosting demonstrated robust performance, especially on larger datasets, while smaller datasets highlighted challenges with models like K-Nearest Neighbors and Naive Bayes due to overfitting and poor generalization. These findings underscore the importance of adapting data-splitting strategies to the characteristics of individual datasets, offering a pathway to improve estimation accuracy in diverse software development scenarios. The research highlights that larger datasets enable better generalization, while smaller datasets require careful model and preprocessing choices to mitigate overfitting and achieve reliable results.