A machine learning model training method and device based on first-order logic rules
By transforming natural language rules validated by domain experts into first-order logical expressions and constructing a system of differentiable logical operators, combined with a method of dynamically adjusting tradeoff coefficients, the problem of insufficient logical constraints in existing technologies is solved, thereby improving the training efficiency and prediction performance of machine learning models in scenarios with low data coverage.
Patent Information
- Application Number
- CN202610843175.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-11
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies lack a complete path to systematically transform domain experience rules into first-order logical expressions that are differentiable, and they also lack a mechanism to dynamically adjust the strength of logical constraints based on data coverage and noise levels, which limits the generalization ability of machine learning models in low data coverage and noisy scenarios.
By transforming a set of natural language rules validated by domain experts into first-order logical expressions, a system of differentiable logical operators is constructed. Combined with a gradient descent optimization framework, the tradeoff coefficients are dynamically adjusted to generate a differentiable rule penalty function. The loss value is then fused for model training until the preset convergence condition is met.
It significantly improves convergence speed and performance in scenarios with simple models and insufficient data coverage, dynamically adjusts the strength of logical constraints, enhances the model's extrapolation ability and prediction stability in data-scarce regions, and achieves the optimal balance between domain knowledge and data signals.
Smart Images
Figure CN122414441A_ABST