The traditional intelligent financial shared service center (IFSSC) faces three challenges: data heterogeneity, task fragmentation and scene dynamics. To address these challenges, this paper proposes a multi-task financial data classification and prediction model (DGAT-MTL) for IFSSC, which aims to establish a multi-task learning paradigm for the financial domain through dynamic weight allocation and domain knowledge injection. The model is based on a dynamic graph attention mechanism and comprises a shared feature encoder, a dynamic task correlation modeling module, a domain knowledge injection module, and a domain adaptation module. By analyzing gradient similarity and quantifying business rules, the model can dynamically adjust task weights, uncover the evolving patterns of financial task correlations, and address domain drift in financial data using Maximum Mean Discrepancy (MMD) and Gradient Reversal Layer (GRL).The experimental results show that DGAT-MTL performs well in the tasks of accounts payable classification and cash flow prediction, with an accuracy rate of 91.5% and a root mean square error (RMSE) of 7.3, and shows strong generalization ability in cross-domain verification. The model not only improves the efficiency and accuracy of financial data processing, but also significantly enhances the stability and adaptability of the model in different business stages and enterprise environments, which provides strong support for the construction of intelligent financial sharing center.
Cai Li (Sun,) studied this question.