<p>Emotion recognition in conversation (ERC) is a crucial research direction in sentiment analysis and human-computer interaction. However, existing models face significant challenges due to two key factors: imbalanced sample distributions across emotion categories and the difficulty of effectively integrating high- and low-frequency information from multimodal signals. To address these challenges, this paper proposes a Curriculum Learning-Guided Frequency Graph Attention Network (CL-FGAN) for ERC. The model employs a Curriculum Learning strategy to address the issue of imbalanced emotion category samples. Additionally, Wavelet Transform technology is utilized to separate high- and low-frequency features in multimodal signals, which are then integrated through a Hybrid Frequency Graph Attention Mechanism to enhance ERC performance and classification accuracy. Extensive experiments demonstrate that the CL-FGAN model outperforms existing representative models on the IEMOCAP and MELD datasets. Specifically, on the MELD dataset, it achieves higher F1 scores for the minority emotion categories of “Fear” and “Disgust”, with improvements of 0.46% and 0.57%, respectively, compared to the baseline model M³Net. These results validate the effectiveness and advantages of the proposed method.</p>