An Unsupervised Multi-View Graph Pseudo-Labeling Framework for Automatic Modulation Recognition

Abstract

Automatic modulation recognition (AMR) in non-cooperative systems is limited by scarce labels. This letter proposes a strictly unsupervised, fully label-free multi-view graph pseudo-labeling framework. Ground-truth labels are never loaded by any training-stage routine: representation learning, graph construction, pseudo-label generation, consensus filtering, hyperparameter setting, checkpointing, and student training use only unlabeled data and unsupervised diagnostics. Labels are accessed solely by an independent evaluation script after configurations, checkpoints, and predictions are frozen. Raw IQ, constellation, FFT, STFT, and expert statistics form two physical spaces whose KNN graphs are fused before strict three-seed consensus filtering; the resulting pseudo-cluster assignments train a raw-IQ student. In the primary intended-configuration comparison with RFFAE-S, SCAN, SKTFAE, and CDC, the method achieves 0.8005 ACC at 18 dB and 0.7538 average ACC over 2–18 dB on RML2016.10A, exceeding the strongest baseline by an absolute ACC margin of 0.0805 at 18 dB. On RML2016.10B, it obtains 0.8193 and 0.7626, with absolute gains of 0.1143 and 0.0892, respectively.