<p>Emotion recognition is a critical technology in the field of affective computing. Among various carriers of emotional information, electroencephalogram (EEG) signals are widely studied due to their ease of acquisition and resistance to deception. However, raw EEG signals suffer from low signal-to-noise ratio and high dimensionality, necessitating effective feature extraction strategies to fully utilize their rich emotional information. Additionally, the directional interactions among brain neurons are not adequately characterized by traditional graph neural network (GNNs), which rely on spectral graph convolution and are limited to undirected graph structures. To address these challenges, this paper proposes a Temporal-Frequency Fusion Multi-Scale Diffusion Convolution Network (TFF-MDCN). First, a temporal-frequency fusion module based on a gating mechanism is developed to dynamically adjust the weights of time and frequency domain features, enhancing EEG signal representation. Second, a graph diffusion convolution mechanism is introduced to overcome the limitations of spectral graph convolution in handling directed graphs, enabling better capture of directional interactions between electrodes. Finally, a self-attention-based multi-scale fusion module is incorporated to capture dependencies across electrodes at varying diffusion steps, facilitating enhanced generalization. Under subject-dependent protocol, TFF-MDCN achieved accuracies of 93.41% on the SEED dataset and 91.21% and 91.85% on the valence and arousal dimensions of the DEAP dataset, respectively. Under subject-independent protocol, the corresponding accuracies were 82.71%, 64.12%, and 65.55%, demonstrating superior performance.</p>