Tree Model Compression by Error-Based Node Pruning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As the number of nodes in a learning tree model increases with improved prediction accuracy, the data storage requirements become substantial, making it challenging to mount the learned algorithm on IC chips with limited storage capacity, while reducing the data amount of the learned model may lower prediction accuracy.
Innovation Solution
A data amount compressing method that reads the error amount associated with each node in the learned model and deletes nodes based on their error amounts, ensuring that nodes with higher errors than their parent nodes are deleted while preserving nodes with high prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of nodes in the learning tree model is increased to improve prediction accuracy, then the prediction accuracy is improved, but the data storage requirements increase making it difficult to mount on IC chips with limited storage capacity
Solution Approach 1:
The patent extracts and removes nodes with high error amounts from the learned model's tree structure. By identifying nodes where the error amount exceeds a predetermined threshold and systematically deleting them, the patent reduces the data storage requirements while attempting to maintain acceptable prediction accuracy levels.
Solution Approach 2:
The patent changes the parameter of node selection criteria from purely structural to error-based. By introducing error amounts as a key parameter for node deletion decisions, the system can selectively remove nodes that contribute least to prediction accuracy, thereby reducing storage requirements in a targeted manner.
2Quantity of substance
If the data amount of the learned model is reduced to fit on IC chips, then the storage capacity requirement is reduced, but the prediction accuracy may be lowered
Solution Approach 1:
The patent introduces error amounts as a new parameter to guide the data reduction process. Instead of arbitrary node removal, nodes are selected for deletion based on their error amounts, ensuring that nodes contributing most to prediction accuracy are preserved while reducing overall data amount.
Solution Approach 2:
The patent applies different deletion criteria to different nodes based on their local error characteristics. Nodes with high error amounts are prioritized for deletion, while nodes with low error amounts are preserved, creating a non-uniform but optimized reduction strategy that maintains local prediction quality where it matters most.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
[Problem to be Solved] The data amount of a learned mode is compressed while minimizing a decrease in prediction accuracy. [Solution] Provided is a data amount compressing method for compressing a data amount corresponding to a learned model obtained by letting the learning model learn a predetermined data group, the learning model having a tree structure in which multiple nodes associated with respective hierarchically divided state spaces are hierarchically arranged, wherein each node in the learned model is associated with an error amount that is generated in the process of the learning and corresponds to prediction accuracy, and the data amount compressing method comprises: a reading step of reading the error amount associated with each node; and a node deleting step of deleting a part of the nodes of the learned model according to the error amount read in the reading step, thereby compressing the data amount corresponding to the learned model.