Different representation learning objectives recover distinct latent structures from the same psychometric data

Psychometric questionnaires are often reduced to aggregate scores, which may obscure granular, item-level information. This study investigates whether different representation learning objectives—specifically contrastive learning versus multi-task learning—recover the same latent structures from the same psychometric data. Using 757 matched teacher-child pairs from the Cyprus ProW preschool trial, the authors compared three approaches: Principal Component Analysis (PCA) for baseline behavioral phenotypes, a dual-encoder contrastive model for teacher-child alignment, and a multi-task model that jointly optimizes for both alignment and behavioral prediction.

The choice of learning objective fundamentally dictates the latent organization of the data. PCA effectively identified four distinct behavioral phenotypes (High Adaptive Functioning to High Behavioral Risk) but provided no information on teacher-child correspondence. Conversely, the contrastive learning framework significantly improved teacher-child retrieval accuracy (Top-10 accuracy increasing from 1.98% to 56.14%) but failed to preserve the behavioral phenotype structure. The multi-task model attempted to bridge this gap by jointly optimizing for both tasks; while it successfully improved behavioral organization, it did so at the cost of significantly reduced teacher-child retrieval performance. This demonstrates a clear trade-off between alignment and behavioral representation.

These results caution researchers against assuming that a single latent representation is universally optimal. In behavioral science, where representation learning is increasingly used to characterize complex populations, the "best" model depends entirely on the scientific goal—whether that is identifying individual behavioral profiles or mapping relationships between teachers and children. The study highlights that latent structures are not inherent properties of the data alone but are co-constructed by the data and the specific objective used to analyze them.