Abstract: Radar signal sorting is a critical function of radar reconnaissance systems for distinguishing radiation sources from complex interleaved pulse streams. Aiming at the problem that only a very limited number of pulses can be labeled and that severe overlap among pulse parameters in highly contested electromagnetic environments leads to degraded sorting accuracy, this paper proposes a semi-supervised learning radar signal sorting method based on a dual-graph residual graph convolutional network. By constructing a feature similarity graph and an anchor graph, the proposed method jointly models the correlations among radar pulses from the perspectives of local feature similarity and global parameter prototype structure. Meanwhile, a multi-subspace parallel residual graph convolutional network is designed to perform parallel modeling of pulse distributions in different feature subspaces, thereby enhancing the representation capability under complex parameter-overlapping scenarios. In addition, a node-wise adaptive gating mechanism is introduced to dynamically adjust the contributions of the two graph branches for different pulse nodes, effectively suppressing the erroneous propagation of low-confidence pseudo-labels over the graph structure. Experimental results demonstrate that, under extremely limited labeled-sample conditions, the proposed method achieves significantly superior sorting performance compared with existing baseline methods, providing an efficient semi-supervised solution for radar signal sorting in complex electromagnetic environments.