This paper presents a subject-independent emotion recognition model, BFE-Net, which leverages EEG frequency band features and self-adaptive graph construction to enhance the robustness and generalizability of emotion recognition. The model aims to address the challenges of individual differences and data heterogeneity in EEG signals. BFE-Net integrates a CNN layer for deep feature extraction, a multi-graphic layer construction module for adaptive graph structure learning, and a GCN layer for feature aggregation. The experimental results on the SEED and SEED-IV datasets demonstrate that BFE-Net outperforms existing methods in terms of accuracy and stability, achieving higher average accuracies and lower standard deviations. The model's effectiveness is further validated through ablation studies and visualization of brain connectivity patterns, confirming its ability to capture complex emotional states and inter-channel relationships.This paper presents a subject-independent emotion recognition model, BFE-Net, which leverages EEG frequency band features and self-adaptive graph construction to enhance the robustness and generalizability of emotion recognition. The model aims to address the challenges of individual differences and data heterogeneity in EEG signals. BFE-Net integrates a CNN layer for deep feature extraction, a multi-graphic layer construction module for adaptive graph structure learning, and a GCN layer for feature aggregation. The experimental results on the SEED and SEED-IV datasets demonstrate that BFE-Net outperforms existing methods in terms of accuracy and stability, achieving higher average accuracies and lower standard deviations. The model's effectiveness is further validated through ablation studies and visualization of brain connectivity patterns, confirming its ability to capture complex emotional states and inter-channel relationships.