KAN surrogates closely reproduce LPA perfectionism profiles in Turkish adolescents

Machine-learning systems trained to reproduce Latent Profile Analysis (LPA) assignments classified perfectionism profiles in 1,385 Turkish adolescents with high fidelity, according to a comparative study in Neural Computing and Applications . The researchers tested Support Vector Machines (SVMs), a neural multilayer perceptron (MLP), Plain Kolmogorov–Arnold Networks (PlainKAN) and an optimized version called EfficientKAN in scenarios based on either four or five profiles. Because the supervised target for every classifier was a profile label derived from the original LPA, the results measure how accurately the surrogates reproduced that statistical solution rather than their independent diagnostic accuracy.

PlainKAN achieved a mean accuracy of 0.947 with a standard deviation of 0.018, while EfficientKAN reached 0.943 ± 0.015. The corresponding results were 0.926 ± 0.013 for the SVM and 0.906 ± 0.025 for the MLP. When the researchers retained only LPA assignments with a classification probability of at least 0.65, PlainKAN rose to 0.992 ± 0.009 and EfficientKAN to 0.989 ± 0.009.

All reported accuracies were means and standard deviations from ten-seed Monte Carlo cross-validation, and the study used pairwise significance tests to compare the algorithms. The five-profile scenarios produced a similar pattern. Without confidence filtering, PlainKAN and EfficientKAN both reached a mean accuracy of 0.927, with standard deviations of 0.015 and 0.020, respectively; the SVM achieved 0.898 ± 0.026.

In the filtered five-profile scenario, EfficientKAN attained 0.987 ± 0.006. The original LPA solution had a mean modal posterior of 0.890, reported as a reference measure of classification certainty rather than as a direct performance estimate for the machine-learning surrogates. The study also examined sensitivity over a classification-probability range of 0.50–0.90.

The practical motivation was rapid classification. LPA identifies latent classes, but the study notes that it cannot classify new cases in real time without re-estimating the entire mixture model. A supervised surrogate instead learns the relationship between measured perfectionism-scale data and the profile labels generated by LPA, allowing it to assign unseen respondents after training.

EfficientKAN achieved fidelity comparable to PlainKAN with roughly 100-fold shorter training time, giving it the strongest reported balance between classification fidelity and computational cost in this comparison. The findings were limited to the study’s sample of Turkish adolescents, its two profile granularities and its filtered and unfiltered analytical scenarios; they do not establish performance in other populations, cultures or clinical settings.

Subject of Research: Machine-learning surrogates for reproducing LPA-derived perfectionism profiles in Turkish adolescents

Article Title: Machine learning surrogates for latent profile analysis: a comparative study with Kolmogorov–Arnold Networks on perfectionism profiles in Turkish adolescents

Article References: Kokcam, O. M., Kokcam, B., & Ucar, F. (2026). Machine learning surrogates for latent profile analysis: a comparative study with Kolmogorov–Arnold Networks on perfectionism profiles in Turkish adolescents. Neural Computing and Applications, 38(20), Article 826. https://doi.org/10.1007/s00521-026-12549-3

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12549-3

Keywords: Kolmogorov–Arnold Networks, Latent Profile Analysis, perfectionism, Turkish adolescents, machine learning, EfficientKAN, classification fidelity, neural networks, Support Vector Machines, Machine, learning, surrogates