Automated Data Hierarchy Pruning for Low-Signal Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional machine learning classifiers trained on data sets with classes having few observations produce inaccurate or uninformative classifications due to low signal-to-noise ratios, leading to misleading data analyses.
Innovation Solution
A system that generates a hierarchical classification by pruning categories with insufficient data items, reassigned to parent categories, and trains a machine learning model using the revised data set to improve classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a traditional classifier is trained on a complete data set including all classes, then the classifier can recognize all possible categories, but the classification accuracy deteriorates due to classes with very few observations producing statistically questionable results
Solution Approach 1:
The patent segments the classification hierarchy by identifying and separating classes with insufficient observations from the main classification structure. This is achieved by calculating a signal-to-noise ratio for each class and removing classes that fall below a threshold, thereby segmenting the problem into manageable parts: well-supported classes and poorly-supported classes.
Solution Approach 2:
The patent extracts and removes classes with very few observations from the training data set. By calculating the signal-to-noise ratio for each class and comparing it against a threshold, the system extracts only the problematic classes (those with insufficient observations) and removes them, preventing these classes from degrading overall classification accuracy.
2Measurement precision
If manual classification rules are used to ensure accurate categorization, then classification precision may be maintained, but the processing productivity deteriorates due to laborious manual classification of millions or billions of data items
Solution Approach 1:
The patent performs preliminary action by pre-calculating the signal-to-noise ratio for each class in the training data set before training the classifier. This preliminary analysis identifies which classes have sufficient observations to support reliable classification, allowing the system to pre-determine the optimal classification structure and avoid manual intervention during actual data processing.
Solution Approach 2:
The system performs self-service by automatically identifying and removing classes with insufficient observations without requiring manual intervention. The classifier autonomously determines which classes to exclude based on statistical criteria (signal-to-noise ratio thresholds), eliminating the need for laborious manual classification while maintaining high processing throughput.
3Quantity of substance
If all classes including those with very few observations are included in the training data, then the data set completeness is maintained, but the signal-to-noise ratio deteriorates leading to misleading classifications
Solution Approach 1:
The patent changes the parameter of class inclusion by dynamically determining which classes should be included in the training data based on their signal-to-noise ratio. Classes with very few observations are identified by calculating their signal-to-noise ratio and are subsequently excluded, thereby changing the composition of the training data to include only classes with sufficient statistical reliability.
Data Source
AI summary
Techniques are disclosed for revising training data used for training a machine learning model to exclude categories that are associated with an insufficient number of data items in the training data set. The system then merges any data items associated with a removed category into a parent category in a hierarchy of classifications. The revised training data set, which includes the recategorized data items and lacks the removed categories, is then used to train a machine learning model in a way that avoids recognizing the removed categories.