<p>Resting-state functional magnetic resonance imaging (rs-fMRI) has shown promise for autism spectrum disorder (ASD) identification, yet cross-site deployment remains challenging due to pronounced site effects, unreliable pseudo-labels, and the insufficient exploitation of functional connectivity (FC) structure. To address these issues, we propose a semantic-aligned Three-Way Decisions (TWD) domain adaptation network (STDAN) for unsupervised cross-site ASD classification. First, we design a structure-aware backbone, which operates directly on FC matrices and captures symmetry-preserving second-order interactions through bilinear subspace projection. Second, a GRL-based adversarial discriminator is introduced to learn domain-invariant embeddings and reduce cross-site distribution discrepancies. Third, a confidence-driven TWD controller partitions target samples into accept/defer/reject regions, enabling risk-aware pseudo-labeling to suppress confirmation bias and error accumulation. Finally, a dynamic class-center memory maintains semantic consistency by updating class prototypes with reliable samples, thereby strengthening class-conditional alignment and stabilizing the decision boundary. Experiments on ABIDE under intra-site and inter-site protocols show that STDAN achieves 73.21% intra-site ACC and 70.32% inter-site ACC on CC200, outperforming representative shallow and deep adaptation baselines. Additional analyses further support its robustness and generalizability for multisite ASD recognition.</p>