Deep + Classical Co-Training Under Scarce Labels for RF Modulation Recognition

We study label-efficiency in RF modulation recognition by co-training a small temporal CNN with a stack of classical
models (RF, SVM, GBM, KNN) using handcrafted features. With
only 0.5% ∼ 10% labels, co-training yields consistent AUROC
gains and improved robustness under test-time SNR shifts. Code
and figures are fully reproducible.