Under the dual background of big data and high-frequency changes in tax policies, corporate tax risk management faces bottlenecks such as data fragmentation, response lag and static rules. This paper proposes an "adaptive tax risk prediction and optimal scheduling algorithm in the enterprise finance and taxation big data environment", which systematically introduces the adaptive mechanism into tax risk governance for the first time. Based on the unified data lake, the algorithm integrates multi-source heterogeneous data such as declaration, invoice, financial report and policy text, and captures policy changes in real time through the dynamic weight layer based on attention mechanism to realize online adjustment of feature importance. The online learning integrated model is used to continuously update the parameters, so that the risk score evolves synchronously with the enterprise operation and macro environment. Furthermore, the risk prediction results are input into the multi-objective optimization scheduling engine, and the Pareto optimal policy set is solved under the dual objectives of "minimum risk-minimum cost" and the hard constraints of cash flow and compliance, so as to generate executable payment, tax refund and declaration adjustment schemes for enterprises. The experiment of backtracking the data of 500 manufacturing enterprises from 2019 to 2023 shows that the AUC of the algorithm is 0.923, and the recall rate is 86.7%, which is 15.5 percentage points higher than that of static XGBoost. Automatically increase the weight of relevant features within 24 hours under the policy surprise scene, and quickly lock the risk point of "fictional R The optimal scheduling of 50 high-risk enterprises saves the compliance cost by 59% on average, and the dispute rate drops to 4%. The research proves that the algorithm can realize the transformation from "experience-driven" to "data-driven" tax management, and provide a new paradigm of real-time, accurate and economical dynamic risk management for enterprises and tax authorities.
Guangqing Deng (Sun,) studied this question.