因子选股系列之一一五：DFQ-diversify：解决分布外泛化问题的自监督领域识别与对抗解耦模型

• The DFQ-Diversify model effectively addresses the out-of-distribution generalization problem by introducing a self-supervised domain recognition and adversarial training mechanism, achieving explicit decoupling of label prediction and domain recognition tasks[2][3][10] - The model's training process includes three core modules: update_d, set_dlabel, and update, which work together through adversarial training to complete domain recognition and label prediction tasks, achieving explicit decoupling of the two[3][22][23] - The update_d module is responsible for domain recognition, using a GRU-based feature extractor, a domain bottleneck layer, a domain classifier, and a label adversarial discriminator to enhance domain representation accuracy and robustness[23][24][25] - The set_dlabel module updates the domain labels of samples through inference and clustering optimization, ensuring that the domain labels reflect the actual distribution of features in the feature space[28][29] - The update module focuses on label prediction, using a shared GRU feature extractor, a label bottleneck layer, a label classifier, and a domain adversarial discriminator to enhance label prediction accuracy and robustness[30][31][32] - The model employs a self-supervised dynamic domain partitioning mechanism, which helps the model autonomously identify potential domain information, enhancing its flexibility and generalization adaptability[34][36] - The DFQ-Diversify model constructs a three-level adversarial training mechanism, including inter-module task adversarial updates, intra-module dual loss adversarial balance, and gradient reversal layer mechanism, to achieve feature decoupling and robust transfer learning[42][43][47] - Compared to the Factorvae-pro model, the DFQ-Diversify model introduces self-supervised learning to dynamically identify potential domains, enhancing flexibility and generalization ability[50][53] - The DFQ-Diversify model shows superior performance in multiple stock pools, especially in large-cap stocks, with significant excess returns in the CSI All Share Index, CSI 300, and CSI 500 stock pools[5][6][107] - The model's backtesting results indicate that it achieved an IC of 12.22%, rankIC of 14.58%, and an annualized excess return of 32.52% in the CSI All Share Index stock pool from 2020 to 2025[5][107] - In the CSI 300 and CSI 500 enhanced strategies, the model achieved an IR of 1.89 and 1.67, and annualized excess returns of 11.27% and 12.19%, respectively[6][172][180]