Forced-choice (FC) test is a widely used approach in noncognitive assessment, which aims to reduce social desirability bias and resist faking. However, existing neural models struggle with FC data due to their lack of design for noncognitive response processes and inadequate comprehensive modeling of hierarchical participant-item relations. To this end, we propose Forced-Choice Relation-based Diagnosis (FCRD), a novel graph neural network that explicitly captures hierarchical participant–item relations. FCRD constructs a unified relational graph integrating participant–item interactions, item–item competitions, and item–item correlations. Furthermore, we design a novel diagnosis layer suited to FC response data, enabling the model to learn from ordinal preference signals and better infer latent noncognitive trait levels. Extensive experiments on real-world datasets demonstrate that FCRD improves diagnostic performance, effectively captures hierarchical relations in FC tests, and provides interpretable insights into its decision process.