Evaluating Kernel Functions in Software Effort Estimation: A Comparative Study of Moving Window and Spectral Clustering Models Across Diverse Datasets

Abstract

This study embarks on an in-depth analysis of the performance of various kernel functions, namely uniform, epanechnikov, triangular, and gaussian, in window-based and spectral clustering-based models. Employing seven distinct datasets, our approach evaluated both window sizes (25%, 50%, 75%, and 100%) and clustering clusters (ranging from 1 to 4). The kernel functions served as weighting functions for regression models, leading to the creation of 192 window-based and 192 clustering-based models. Our analysis underscores the dominance of the uniform kernel function. In most models where the Pred(0.25) was maximal and the Mean Absolute Error was minimal, the uniform kernel function was predominantly utilized. Further, our results exhibit varying outcomes between moving windows and spectral clustering across datasets. For instance, in the fpa_china dataset, while moving windows with a 50% size displayed no significant superiority over spectral-clustering with 1 cluster, spectral-clustering (1 cluster) demonstrated a significantly enhanced performance. However, in datasets like fpa_kitchenham, neither approach proved to be significantly better. This comprehensive exploration into the efficiency of kernel functions in moving windows and spectral-clustering models provides valuable insights for future research and applications in data modelling and analysis.INDEX TERMS software effort estimation, kernel function, moving windows, spectral clustering, functional points, use case points