KAN‐SEE (Kolmogorov–Arnold Networks for Software Effort Estimation): A Function‐Based Extension of COCOMO

ABSTRACT

Accurate estimation of software development effort remains a persistent challenge in software engineering. Traditional parametric models such as COCOMO provide structured estimation mechanisms but often fail to capture the nonlinear interactions among project cost drivers. Machine learning approaches improve predictive accuracy but typically sacrifice the interpretability that practitioners require for managerial decision‐making. Recent advances in learning architectures, specifically Kolmogorov–Arnold Networks (KAN), offer a function‐based modelling paradigm that learns structured, data‐adaptive nonlinear relationships while preserving symbolic interpretability through post‐training spline analysis. This paper proposes KAN‐SEE , a KAN‐based software effort estimation framework that maps four heterogeneous benchmark datasets—COCOMO81, China, NASA93, and ISBSG—into a unified 16‐dimensional COCOMO feature space consisting of 15 standard effort multipliers and one log‐transformed size feature. The proposed model is evaluated on 7015 software projects using fivefold cross‐validation. Experimental results demonstrate that KAN‐SEE achieves a mean of 0.9652 ( 0.0094) in log‐transformed space.These findings suggest that function‐learning architectures represent a principled and practically viable direction for advancing effort estimation: combining the interpretability of COCOMO with the predictive flexibility of modern machine learning.