Decision Tree Score Information for Feature Range Specification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for risk assessment of information leakage from learning models require statistical data, making assessments impossible without it, and probabilistic feature estimation in average cases may not provide appropriate results, especially for decision trees.
Innovation Solution
A processing device and method that acquire score information from a decision tree to specify the possible range of an unknown feature's value, allowing for risk assessment even without statistical data by using confidence scores from the decision tree.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If statistical data of learned data is used for risk assessment, then assessment can be performed with existing methods, but assessment becomes impossible when statistical data is not available
Solution Approach 1:
The patent creates pseudo-data by copying the statistical characteristics of the original learned data. A second learning model is trained on this generated pseudo-data, allowing risk assessment to proceed even when original statistical data is unavailable. The pseudo-data replicates the essential patterns and distributions needed for meaningful assessment.
Solution Approach 2:
The patent performs preliminary learning to generate pseudo-data before conducting the risk assessment. By pre-training a second learning model on synthesized data that mirrors the original data's characteristics, the system prepares the necessary assessment foundation in advance, enabling evaluation without direct access to original statistical data.
2Ease of operation
If probabilistic feature estimation in average case is performed, then general estimation can be obtained, but appropriate assessment may not be achieved depending on output status
Solution Approach 1:
The patent transitions from average-case probabilistic estimation to local-quality assessment by examining specific nodes in the decision tree. Score information representing the number of training data pieces that fell to each node enables precise, location-specific evaluation rather than generalized probability estimates, improving accuracy for particular output statuses.
Solution Approach 2:
The patent replaces the probabilistic statistical mechanism with a deterministic counting mechanism. Instead of using probability distributions and average-case analysis, the system directly counts the number of training data pieces associated with each decision tree node, providing exact local frequency information that substitutes for approximate probabilistic estimates.
3Measurement precision
If decision tree node data distribution is analyzed, then precise local assessment is achieved, but complexity of processing increases
Solution Approach 1:
The decision tree structure inherently contains the score information needed for local assessment within its own nodes. The patent leverages this self-contained data distribution information, where each node already stores the count of training data pieces that reached it. This eliminates the need for external statistical data or complex post-processing, as the tree serves its own assessment needs.
Data Source
AI summary
A processing device includes an acquisition unit and a specifying unit. The acquisition unit acquires, from a decision tree that is a learned model and includes a plurality of nodes, score information representing a value according to the number of pieces of data that fell to each of the nodes, among a plurality of pieces of training data used for training of the decision tree. The specifying unit specifies a possible range that the value of an unknown feature may take, on the basis of the score information acquired by the acquisition unit. The unknown feature is a part of the features included in the training data.