Decision Tree Training Data Influence Evaluation With Similarity Scoring

The patent segments the training data evaluation process into two distinct phases: (1) a similarity score calculation phase that computes similarity between each training data point and others using the trained model's tree structure, and (2) an influence score calculation phase that only evaluates target data selected based on similarity thresholds. This segmentation avoids the need to learn reference models for all training data while maintaining evaluation accuracy for the most influential data points.

The patent applies partial action by selecting only a subset of training data (target data) for influence score evaluation based on similarity scores, rather than evaluating all training data. By setting a similarity threshold, the system focuses computational resources on evaluating only those data points that are most similar to the query data, thereby reducing processing time while maintaining evaluation accuracy for critical cases.

The patent creates a universal evaluation method that works across different machine learning model types by leveraging the common concept of similarity in feature space. The similarity score calculation using the trained model's structure (tree structure for decision trees, or other model structures) provides a model-agnostic approach that can be applied to decision trees, deep learning models, and other types, thereby achieving both speed and versatility.

The patent introduces similarity score as an intermediary metric that bridges the gap between different model types. Instead of directly using model-specific characteristics (which would limit applicability), the similarity score serves as a universal mediator that captures the essential relationship between training data points in a way that is applicable across different model architectures while still enabling efficient evaluation.