Decision Tree Node Data Pre-Statistics Mechanism
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional decision tree generation processes require frequent access to large datasets, leading to significant processing delays and reduced efficiency, especially in applications with massive data samples and features.
Innovation Solution
A data processing mechanism that performs pre-statistics on feature values to determine a target range with a high number of feature values, allowing only the remaining values to be assigned to numerical ranges, thereby reducing the number of data accesses and improving processing speed and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional decision tree generation processes are used, then accurate classification and regression can be achieved, but processing time increases significantly due to frequent accesses to large datasets
Solution Approach 1:
The patent applies preliminary action by performing pre-statistics on feature values before the actual decision tree generation process. Histograms of feature values are pre-calculated and stored, allowing the algorithm to quickly determine optimal split points without repeatedly accessing the full training dataset during tree construction. This pre-processing step significantly reduces processing time while maintaining classification accuracy.
Solution Approach 2:
The patent extracts only the necessary statistical information (histogram data) from the training dataset rather than accessing the complete dataset repeatedly. By extracting and storing pre-computed feature value distributions, the method eliminates the need for frequent full-data accesses during decision tree generation, thereby reducing processing time while preserving the ability to make accurate split decisions.
2Manufacturing precision
If frequent accesses to training data are performed, then accurate node splitting can be achieved, but processing efficiency substantially reduces
Solution Approach 1:
The patent performs preliminary computation of feature value histograms and stores them for rapid access during node splitting. This pre-computed statistical information enables accurate determination of optimal split points without requiring repeated scans of the training data, thus maintaining node splitting accuracy while dramatically improving processing efficiency.
Solution Approach 2:
The patent creates a compressed representation (copy) of the training data in the form of histograms that capture the essential distribution characteristics. These histogram copies are stored and reused during decision tree construction, eliminating the need for frequent accesses to the original large training dataset while preserving the information needed for accurate node splitting decisions.
3Loss of information
If all feature values are assigned to numerical ranges, then complete data distribution information is obtained, but data access and storage requirements increase
Solution Approach 1:
The patent segments the continuous feature value space into discrete numerical ranges (bins) and computes histograms that count the distribution of values within each range. This segmentation transforms the problem from handling individual feature values to managing aggregated range statistics, significantly reducing data access and storage requirements while preserving the essential distribution information needed for decision tree splitting.
Solution Approach 2:
The patent changes the parameter representation from individual feature values to aggregated histogram counts across numerical ranges. By transforming the data from fine-grained individual values to coarser-grained range distributions, the method reduces the volume of data that needs to be accessed and stored while maintaining sufficient information for accurate decision tree construction.
Data Source
AI summary
Embodiments of the subject matter described herein relate to generating a decision tree based on data pre-statistics. A plurality of data samples for a node of the decision tree are obtained, and the plurality of data samples have corresponding feature values with respect to a first feature. A target range is determined from a plurality of predefined numerical ranges so that the number of feature values falling into the target range is greater than a predetermined threshold number. Then, the remaining of the feature values other than the feature values falling into the target range are assigned to the respective numerical ranges, and the feature values falling into all the numerical ranges are counted based on the assignment of the remaining of the feature values, for allocation of the plurality of data samples to child nodes of the node. Accordingly, the data processing efficiency is substantially improved.