Effectively mining emotional states from multi-channel electroencephalography (EEG) signals presents significant data mining challenges. Prevailing methods often limit recognition accuracy by treating signals as independent time-series. This approach neglects crucial inter-channel correlations and results in non-interpretable ‘black-box’ models. To overcome these deficiencies, we propose the Relational Probabilistic Graph Convolutional Network (RPGCN), a novel two-stage framework. First, RPGCN transforms raw EEG signals into probabilistic graphs that explicitly model dynamic, inter-channel dependencies to capture latent emotional variations. Subsequently, a Graph Convolutional Network (GCN) learns discriminative patterns directly from these structured representations for classification. Extensive experiments demonstrate that RPGCN significantly outperforms state-of-the-art methods. Crucially, our framework provides inherent model interpretability, with learned relational patterns aligning with established cognitive neuroscience findings. Our work establishes a new, interpretable graph-based paradigm for EEG analysis, paving the way for more robust and intelligent human-computer interaction.