Hierarchical Classifier Using Divergence-Based Splitting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing classification methods, such as greedy algorithms, often converge to local maxima or plateaus, failing to achieve a globally optimal solution, especially in multi-modal datasets, leading to increased classification errors.
Innovation Solution
A hierarchical classifier is constructed using a divergence measure to determine the type of node, with divergence-based splitting to maximize class separability and information-based splitting to minimize misclassification error, allowing for both classification and clustering nodes to reduce complexity and error.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If greedy algorithms are used for classification, then training error decreases locally at each node, but the system converges to local maxima and fails to achieve globally optimal solution
Solution Approach 1:
The patent segments the classification problem into hierarchical levels, where each level divides data into subsets. This segmentation allows the system to explore multiple local optima at different levels and combine them to achieve a globally optimal solution, avoiding the limitation of single-level greedy algorithms that get trapped in local maxima.
Solution Approach 2:
The patent introduces a hierarchical dimension to the classification process. Instead of optimizing a single classification tree, it builds multiple levels of hierarchy where each level refines the classification. This dimensional expansion allows the system to escape local optima by exploring solutions across multiple hierarchical layers.
2Measurement precision
If multiple splitting criteria are used to maximize class separability and minimize misclassification error, then classification accuracy improves, but computational complexity increases
Solution Approach 1:
The patent dynamically selects between divergence-based splitting criterion and information-based splitting criterion based on the current node's characteristics. This dynamic adaptation allows the system to use the most appropriate criterion for each specific situation, improving classification accuracy while avoiding the unnecessary computational overhead of always using both criteria.
Solution Approach 2:
The patent applies different splitting criteria to different parts of the decision tree based on local data characteristics. Divergence-based splitting is used when class separability is the primary concern, while information-based splitting is used when minimizing misclassification error is more important. This localized application of different criteria optimizes the balance between accuracy and complexity.
Data Source
AI summary
Described herein is a framework for constructing a hierarchical classifier for facilitating classification of digitized data. In one implementation, a divergence measure of a node of the hierarchical classifier is determined. Data at the node is divided into at least two child nodes based on a splitting criterion to form at least a portion of the hierarchical classifier. The splitting criterion is selected based on the divergence measure. If the divergence measure is less than a predetermined threshold value, the splitting criterion comprises a divergence-based splitting criterion which maximizes subsequent divergence after a split. Otherwise, the splitting criterion comprises an information-based splitting criterion which seeks to minimize subsequent misclassification error after the split.