Method, device, equipment and medium for extracting formula-performance statistical correlation
Patent Information
- Application Number
- CN202611320665.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-10-16
AI Technical Summary
[0003]传统配方-性能建模方法主要包括：基于人工特征的统计回归模型，通常仅能预测单一或少数性能指标；基于向量化输入的机器学习模型，将多个性能指标作为多输出回归任务，但缺乏对标签空间结构的建模
[0009]The beneficial effects of this invention are as follows: The method for extracting the statistical correlation between formulation and performance provided by this invention first extracts formulation graph embedding features and test condition features from the target formulation data, which can encode the formulation component information into a graph structure representation and simultaneously acquire test condition information; then, the formulation graph embedding features are mapped to initial performance features and label space message passing is performed based on an adaptive label graph structure, which can refine the initial performance features in the label space using a label graph structure dynamically constructed in advance based on the statistical correlation between different performance indicators in large-scale data; then, the formulation graph embedding features, test condition features, and refined label graph features are used as the initial representations of formulation nodes, condition nodes, and performance nodes in the ternary graph structure, and message passing is performed within the ternary graph structure, which can realize the ternary interaction between formulation information, condition information, and performance information, so that the condition information can correct the extraction process of performance features; finally, based on the corrected performance features, the loss of multiple prediction tasks is adaptively weighted and balanced, and the statistical correlation between formulation and performance is output, which can automatically balance the weights of each prediction task in a multi-task learning framework without relying on manual experience to set the weights of each task. This invention can extract statistical correlations between formulation and performance on large-scale data without explicit modeling of process conditions, thereby improving the data processing efficiency and applicability of related computing devices in scenarios where process conditions are missing or extremely complex.