Network alignment in psychometrics: comparing symptom networks in eating disorders across age and diagnosis

Network alignment (NA) is an emerging tool in network science, enabling knowledge transferring and structural comparison of networks. While widely used in biology, social networks and computer vision, its application to network psychometrics has not been explored. In this study we present a comparative analysis of several NA algorithms, spanning four methodological categories: spectral methods, representation learning, deep matching, and genetic algorithms. These algorithms were evaluated using psychometric networks derived from the Eating Disorder Inventory-3 (EDI-3) dataset: in these networks, nodes represent symptoms/personality traits related to eating disorders, while edges encode for statistical dependencies between nodes. Initially, we tested each algorithm’s robustness to structural noise by aligning a psychometric network with permuted and noise-corrupted versions of itself. The best-performing algorithm from each category was then used to align networks representing different age groups (i.e., early vs. late adolescence; early adolescence vs. adulthood) and diagnostic categories (anorexia nervosa vs. bulimia nervosa). The overall best-performing algorithm was subsequently run 100 times to identify the most consistently aligned node pairs, which were interpreted from a psychological perspective. Our results show that PALE, a representation learning–based algorithm, demonstrated the highest robustness and stability under noisy conditions and consistently outperformed others in real-world applications. Notably, PALE identified symptom/personality trait nodes with consistent functional roles across developmental stages, such as interpersonal difficulties and body-related concerns, highlighting their persistent influence within the symptom networks of patients with eating disorders.