Machine learning surrogates for latent profile analysis: a comparative study with Kolmogorov–Arnold Networks on perfectionism profiles in Turkish adolescents
Neural Computing and Applications, cilt.38, sa.20, 2026 (Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 38 Sayı: 20
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s00521-026-12549-3
- Dergi Adı: Neural Computing and Applications
- Derginin Tarandığı İndeksler: Scopus, Aerospace Database, Applied Science & Technology Source, Compendex, Index Islamicus, INSPEC, zbMATH, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Technology Collection (ProQuest)
- Anahtar Kelimeler: Adolescent psychometric assessment, Kolmogorov–Arnold Networks (KAN), Latent profile analysis (LPA), Machine learning (ML), Perfectionism, Surrogate modelling
- Uşak Üniversitesi Adresli: Evet
Özet
Distinguishing adaptive and maladaptive perfectionism among youth is a crucial task for mental health professionals. Latent Profile Analysis (LPA) is widely used for this modeling; however, it cannot classify new cases in real time without re-estimating the whole mixture model. To this end, we develop fast surrogate classifiers that emulate LPA-derived profile assignments for unseen respondents. Since the supervised target is itself an LPA-derived label, reported metrics quantify fidelity rather than independent diagnostic accuracy. On 1385 Turkish adolescents assessed with two multidimensional perfectionism scales, we benchmark classical algorithms, including Support Vector Machines (SVM), together with a neural Multilayer Perceptron (MLP) baseline, against Kolmogorov–Arnold Networks (PlainKAN) and the optimized EfficientKAN. Four scenarios cross four- and five-profile granularity with a CPROB label-confidence filter, with sensitivity verified over a 0.50–0.90 grid. All accuracies are ten-seed Monte Carlo cross-validation means ± SD, backed by pairwise significance tests. On the unfiltered four-class scenario (S1), the KAN variants lead with and, above SVM () and MLP (). On the filtered four-class scenario (S2), PlainKAN reaches and EfficientKAN. On the unfiltered five-class scenario (S3), the KAN variants tie at and, ahead of SVM (). On the filtered five-class scenario (S4), EfficientKAN attains ; LPA’s mean modal posterior of 0.890 is reported as a reference measure of classification certainty in the original solution. EfficientKAN matches PlainKAN's fidelity at a roughly 100-fold shorter training time; hence, it offers the best fidelity–cost balance.