Orsoni Matteo, Giovagnoli Sara, Garofalo Sara, Mazzoni Noemi, Spinoso Matilde, Benassi Mariagrazia
Department of Psychology, University of Bologna, Bologna, Italy.
Front Psychol. 2025 May 9;16:1474292. doi: 10.3389/fpsyg.2025.1474292. eCollection 2025.
Understanding individual cognitive profiles is crucial for developing personalized educational interventions, as cognitive differences can significantly impact how students learn. While traditional methods like factor mixture modeling (FMM) have proven robust for identifying latent cognitive structures, recent advancements in deep learning may offer the potential to capture more intricate and complex cognitive patterns.
This study compares FMM (specifically, FMM-1 and FMM-2 models using age as a covariate) with a Conditional Gaussian Mixture Variational Autoencoder (CGMVAE). The comparison utilizes six cognitive dimensions obtained from the PROFFILO assessment game.
The FMM-1 model, identified as the superior FMM solution, yielded two well-separated clusters (Silhouette score = 0.959). These clusters represent distinct average cognitive levels, with age significantly predicting class membership. In contrast, the CGMVAE identified ten more nuanced cognitive profiles, exhibiting clear developmental trajectories across different age groups. Notably, one dominant cluster (Cluster 9) showed an increase in representation from 44 to 54% with advancing age, indicating a normative developmental pattern. Other clusters displayed diverse profiles, ranging from subtle domain-specific strengths to atypical profiles characterized by significant deficits balanced by compensatory abilities.
These findings highlight a trade-off between the methodologies. FMM provides clear, interpretable groupings suitable for broad classification purposes. Conversely, CGMVAE reveals subtle, non-linear variations in cognitive profiles, potentially reflecting complex developmental pathways. Despite practical challenges associated with CGMVAE's complexity and potential cluster overlap, its capacity to uncover nuanced cognitive patterns demonstrates significant promise for informing the development of highly tailored educational strategies.
了解个体的认知概况对于制定个性化教育干预措施至关重要，因为认知差异会对学生的学习方式产生重大影响。虽然像因子混合模型（FMM）这样的传统方法已被证明在识别潜在认知结构方面很强大，但深度学习的最新进展可能提供捕捉更复杂和精细认知模式的潜力。
本研究将FMM（具体而言，使用年龄作为协变量的FMM - 1和FMM - 2模型）与条件高斯混合变分自编码器（CGMVAE）进行比较。比较使用从PROFFILO评估游戏中获得的六个认知维度。
被确定为最佳FMM解决方案的FMM - 1模型产生了两个明显分开的聚类（轮廓系数= 0.959）。这些聚类代表了不同的平均认知水平，年龄显著预测类别归属。相比之下，CGMVAE识别出了十种更细微的认知概况，在不同年龄组中呈现出清晰的发展轨迹。值得注意的是，一个主要聚类（聚类9）显示随着年龄增长，其占比从44%增加到54%，表明存在一种规范的发展模式。其他聚类呈现出多样的概况，从细微的特定领域优势到以显著缺陷与补偿能力相平衡为特征的非典型概况。
这些发现突出了两种方法之间的权衡。FMM提供了适用于广泛分类目的的清晰、可解释的分组。相反，CGMVAE揭示了认知概况中的细微、非线性变化，可能反映了复杂的发展路径。尽管与CGMVAE的复杂性和潜在聚类重叠相关存在实际挑战，但其揭示细微认知模式的能力为制定高度定制的教育策略提供了重大前景。
