Adaptive heuristic method for constructing decision trees based on data distribution and feature importance
AssigneeLeptude, Inc.
InventorJohn A. Fortkort
Life of this patent
Estimating the term…
Abstract
A method is provided for constructing decision trees. The method includes analyzing a dataset to determine data distribution metrics; calculating feature importance scores for features in the dataset; dynamically adjusting tree depth and node pruning criteria based on the determined data distribution metrics and feature importance scores; initializing a decision tree structure based on the adjusted tree depth and node pruning criteria; and iteratively refining the decision tree by applying heuristic adjustments to improve splits based on updated data distribution metrics and feature importance scores.
Claims20
A method for guiding heuristic initialization in decision tree learning, comprising:
- collecting performance data on decision tree nodes during an initial tree construction phase;
- training a meta-learning model using the collected performance data, wherein the meta-learning model is configured to predict the potential benefit of SAT-based refinement for different parts of the decision tree;
- evaluating nodes or subtrees of the decision tree using the trained meta-learning model to predict which parts of the decision tree are likely to benefit most from SAT-based refinement;
- prioritizing nodes or subtrees for SAT-based refinement based on the predictions of the meta-learning model;
- and
- refining the prioritized nodes or subtrees using SAT-based methods.
Dependencies are inferred from the claim text.