Rule-Based Event Classifier Induction Under Low-FPR Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing rule systems require significant domain expert knowledge for creation and updating, and their predictive performance is often worse than modern ML models, especially in high-stake tasks with extreme class imbalance and low False Positive Rate (FPR) requirements, which is problematic for manufacturing and service operations.
Innovation Solution
A computer-implemented method using GPU or NPU parallel processing for automatic rule generation, incorporating Fast Interpretable Greedy-Tree Unions (FIGU) to dynamically generate and update rule-based classifiers with low FPR, by extracting rules from decision trees and iteratively selecting rules with high precision, while maintaining interpretability and reducing complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If rule systems are manually created and tuned by domain experts, then interpretability is maintained, but predictive performance is worse than modern ML models
Solution Approach 1:
The system enables automatic rule generation from data without requiring manual expert intervention. The rule induction algorithm autonomously extracts rules from training data, allowing the system to self-improve its predictive performance while maintaining interpretability through generated rules.
Solution Approach 2:
The patent segments the complex task of rule generation into distinct phases: data preprocessing, rule induction from training data, rule evaluation, and rule selection. This segmentation allows automated processing while maintaining manageable complexity at each stage.
2Extent of automation
If induction algorithms are used to infer rules from data, then automatic rule generation is achieved, but handling extreme class imbalance with low FPR requirements becomes problematic
Solution Approach 1:
The system changes key parameters during rule induction including sampling strategies for imbalanced classes, threshold adjustments for rule evaluation, and weighted metrics that prioritize low false positive rates. These parameter adjustments enable effective handling of extreme class imbalance while maintaining automatic generation.
Solution Approach 2:
The patent implements feedback mechanisms where generated rules are evaluated against validation data with specific focus on FPR metrics. Poor performing rules are refined or discarded based on this feedback, iteratively improving the system's ability to handle imbalanced classes with low FPR requirements.
3Measurement precision
If complex ML models are used to improve predictive performance, then accuracy increases, but interpretability is lost
Solution Approach 1:
The system creates simplified copies of complex ML model behavior in the form of human-readable rules. Instead of using complex black-box models, the patent extracts rule representations that copy the decision-making logic in an interpretable format, maintaining both performance and understandability.
Solution Approach 2:
The generated rule system serves multiple functions: it provides predictive performance comparable to complex ML models, maintains full interpretability for domain experts, and enables easy modification and updating of rules. This multi-functionality resolves the trade-off between accuracy and interpretability.
4Adaptability or versatility
If rules are manually updated by experts to keep up with new patterns, then adaptability is maintained, but significant domain expert knowledge is required
Solution Approach 1:
The system performs self-updating by automatically inducing new rules from incoming data or updated training sets. When new patterns emerge, the rule induction algorithm autonomously generates updated rules without requiring expert intervention, maintaining adaptability while eliminating the need for continuous expert involvement.
Solution Approach 2:
The patent prepares the system for future updates by establishing automated rule induction capabilities in advance. This preliminary automation ensures that when new patterns need to be detected, the system can immediately begin automatic rule generation without waiting for expert analysis, enhancing adaptability while reducing complexity.
Data Source
Figure 1~3
AI summary
Computer-implemented method for generating and updating a rule-based classifier, for verification of data records of a dataset, each said data record corresponding to an event of a computer system, from an initial ruleset stored in a non-volatile memory of a data verification computer system, the method comprising: receiving a training dataset wherein the training dataset comprises a plurality of data records and, if existing, corresponding classification labels; splitting the training dataset into an induction dataset and a selection dataset; applying the initial ruleset to each of the plurality of data records of the induction dataset for classifying, when applicable, the data records; automatically generating a candidate ruleset from non-classified and labelled records; and applying the candidate ruleset to the selection dataset and iteratively selecting a rule from the candidate ruleset that provides the highest classification precision to the selection dataset, until a stopping criterium or criteria is reached.