Exploring Model-Dependent Learning Profiles in PIRLS 2021: A Comparative Study of Classical and Deep Clustering Methods

Abstract

This study investigates the use of deep clustering to explore latent learning profiles within the PIRLS 2021 dataset, focusing specifically on Moroccan fourth–grade students. Rather than assuming the presence of well-defined groups in the raw data, deep clustering is used as a representation learning approach that constructs a latent space where educational patterns may become more interpretable. By jointly learning embeddings and cluster assignments, these models aim to overcome some of the limitations of classical algorithms such as k-means, agglomerative clustering, and Gaussian Mixture Models when applied to large, heterogeneous educational datasets. While the separation between clusters remains limited in the original feature space, the latent representations learned by deep models allow for the emergence of more coherent profiles that reflect meaningful combinations of linguistic, socio-economic, and school-level factors. These findings suggest that deep clustering can serve as a useful analytical tool for characterizing complex learner populations and generating data-driven insights into the determinants of reading performance in the Moroccan context. This study critically compares classical and deep clustering methods for constructing exploratory learning profiles from the Moroccan PIRLS 2021 data. DEC and IDEC produce sharply separated partitions in their learned latent spaces, but the same assignments remain weakly separated in the common standardized input space (e.g., Silhouette = 0.0221 for DEC). The latent-space scores are therefore interpreted as properties of an optimized representation, not as evidence of naturally occurring or ground-truth student categories. To examine utility beyond internal geometry, we use held-out PIRLS reading plausible values and educational variables that were not used to form the clusters. Across random seeds, cluster membership explains approximately 3.3% − 7.5% of the weighted variation in reading achievement, while home resources show the clearest external contrast. However, assignment stability is low (mean adjusted Rand index = 0.181 for DEC and 0.165 for IDEC), and assignments are also sensitive to missing-data treatment (type-aware versus iterative-imputation ARI = 0.191-0.423 across methods). Thus, deep clustering offers a potentially useful exploratory segmentation of educational gradients. However, a matched 16- component PCA+KMeans baseline is substantially more stable (mean ARI = 0.702) and exceeds the deep methods on all three common-space indices. The evidence therefore does not establish either a unique taxonomy of Moroccan learners or the necessity of a complex deep clustering objective. Educationally, the analysis does not claim to discover new determinants; it refines established relationships by showing that home resources discriminate the fitted profiles more strongly than home-language frequency and that these familiar gradients do not form a unique, stable set of learner types. A harmonized replication on France, Singapore, and South Africa reproduces the main cross-space result: DEC’s latent Silhouette remains 0.9991-0.9997, whereas its common-space Silhouette is only 0.0093-0.0571. Across the four countries, KMeans and PCA+KMeans have higher mean common-space Silhouette and external educational effects than the deep methods. Thus, the methodological warning about latent geometry is cross-country robust, but the magnitude and educational meaning of the profiles remain country-dependent.