Tree Split Criterion Selection Using Representative Data Samples
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning techniques for determining tree structure division criteria require processing loads due to referencing all or most data pieces, leading to slow learning speeds and potential inappropriate divisions, especially when data is scarce.
Innovation Solution
An information processing apparatus that generates division criterion candidates from a plurality of arbitrarily selected data pieces, evaluating and determining an optimal criterion without referencing all data, using methods like average values or histograms, to facilitate rapid and appropriate data division.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If exhaustive search is performed using all data pieces to determine division criterion, then division accuracy is improved, but processing time increases
Solution Approach 1:
The patent segments the data pieces into multiple groups and selects representative data pieces from each group to serve as division criterion candidates. This segmentation approach reduces the number of candidates that need to be evaluated while still maintaining accurate division criterion determination, thereby resolving the contradiction between accuracy and processing time.
Solution Approach 2:
The patent extracts representative data pieces from the full dataset to form a reduced candidate set. By taking out only the essential representative samples rather than processing all data pieces, the system achieves accurate division criterion selection with significantly reduced processing time.
2Reliability
If normalization is performed on all data pieces, then division criterion reliability is improved, but calculation load increases
Solution Approach 1:
The patent segments the data into multiple groups and performs normalization only on representative samples from each segment rather than on all data pieces. This reduces the calculation load while maintaining reliability through the representative nature of the segmented samples.
Solution Approach 2:
The patent uses lightweight, approximate normalization on a small set of representative data pieces rather than exhaustive normalization on all data. This disposable approach on reduced data achieves sufficient reliability with much lower computational cost.
3Measurement precision
If histogram creation is performed on all data pieces, then division criterion appropriateness is improved, but processing complexity increases
Solution Approach 1:
The patent segments the data and creates histograms only from representative samples of each segment rather than from all data pieces. This segmentation reduces processing complexity while maintaining the appropriateness of division criteria through the representative characterization of data distribution.
Solution Approach 2:
The patent extracts key statistical characteristics from representative data pieces to form simplified histograms that capture essential data distribution patterns. This extraction approach achieves appropriate division criterion determination with reduced processing complexity.
Data Source
AI summary
An information processing apparatus includes a candidate generator configured to generate a plurality of data division criterion candidates by generating data division criterion candidates on the basis of a plurality of data pieces arbitrarily selected from the data pieces to be divided at nodes that constitute the tree structure and hold the data pieces to be divided, data division processor circuitry configured to divide the data pieces to be divided on the basis of the plurality of data division criterion candidates to generate a plurality of data division results, evaluation processor circuitry configured to evaluate the data division results to respectively generate evaluation results, and division criterion determination processor circuitry configured to determine one data division criterion candidate among the plurality of data division criterion candidates as a data division criterion on the basis of the evaluation results.