Rule-Based Classifier Updating for Low-FPR Event Verification
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 computationally resource-intensive and not well-aligned with current computing hardware capabilities.
Innovation Solution
A computer-implemented method for generating and updating rule-based classifiers using GPU or NPU parallel processing, incorporating a modified Greedy-Tree Sums (FIGU) algorithm to automatically generate rules with low FPR, involving decision tree generation and iterative rule selection to maintain or improve performance while reducing complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual rule creation and tuning is used, then expert domain knowledge is required, but predictive performance is worse than modern ML models and requires significant expert input
Solution Approach 1:
The system automatically generates and updates rules by inducing them from data patterns rather than requiring manual expert creation. The rule induction algorithm autonomously analyzes data, identifies patterns, and creates rules that meet FPR constraints without human intervention in the rule formulation process.
Solution Approach 2:
The patent replaces manual expert knowledge entry with an automated computational system. Instead of experts manually creating rules, the system uses machine learning techniques to automatically induce rules from data, substituting the mechanical process of manual rule creation with automated algorithmic rule generation.
2Measurement precision
If traditional rule induction algorithms are used, then rules can be inferred from data, but they struggle with extreme class imbalance and require very low FPR values under 2%
Solution Approach 1:
The system dynamically adjusts parameters during rule induction to handle extreme class imbalance. By modifying the induction algorithm to account for imbalanced data distributions and enforcing hard FPR constraints through parameter adjustments, the system achieves both low FPR performance and adaptability to extreme class imbalance conditions.
3Ease of operation
If rules are manually created and tuned, then experts can review and understand decisions, but the process is time-consuming and requires continuous manual modification to keep up with new patterns
Solution Approach 1:
The system continuously automatically updates rules as new data patterns emerge, eliminating the need for periodic manual revisions. The rule induction algorithm operates continuously or at regular intervals to maintain rules up-to-date with current data distributions, providing continuous adaptation without manual intervention.
Solution Approach 2:
The system automatically maintains and updates rules without requiring expert time for manual modification. Experts can still review rules for interpretability, but the bulk of rule maintenance is handled automatically by the induction algorithm that continuously adapts to new patterns in the data.
4Ease of manufacture
If separate-and-conquer algorithms are used, then rule sets can be formed by adding rules one by one, but the process is computationally intensive and not well-aligned with current computing hardware
Solution Approach 1:
The patent segments the rule formation process into independent, parallelizable tasks. By dividing the data into chunks and processing rules in parallel across multiple computing units, the system maintains the systematic approach of separate-and-conquer while achieving computational efficiency through parallelization that aligns with modern hardware capabilities.
Data Source
AI summary
Generating and updating a rule-based classifier, for verification of data records of a dataset, each 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 by: 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.