Rule Set Induction for Redundant Rule Pruning
Solution Overview
Problem
Current rule induction methods are inefficient in handling large datasets and often result in redundant rules, leading to suboptimal prediction accuracy and complexity in interpreting the decision-making process.
Innovation Solution
The proposed method involves creating a rule set hierarchy, where rules are distributed based on consequent labels, and redundant rules are deleted if the ratio of co-occurrence with another rule satisfies a threshold, optimizing the rule set and improving prediction efficiency while allowing for transparent rule inspection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If rule induction is performed on large datasets using traditional methods, then comprehensive pattern coverage is achieved, but computational efficiency deteriorates and processing time increases
Solution Approach 1:
The patent segments the rule induction process into multiple phases: initial rule generation, rule distribution into partitions based on consequent labels, and iterative pruning. This segmentation allows the system to handle large datasets by processing rules in manageable chunks rather than all at once, improving computational efficiency while maintaining comprehensive pattern coverage.
Solution Approach 2:
The patent performs preliminary actions by distributing rules into partitions based on consequent labels before the full induction process completes. This preliminary organization enables efficient pruning decisions to be made early in the process, reducing the computational burden of subsequent processing steps while ensuring all patterns are considered.
2Reliability
If traditional rule induction methods are used, then all potential rules are generated, but rule set size increases and redundancy increases
Solution Approach 1:
The patent extracts and removes redundant rules from the rule set through a systematic pruning process. By calculating co-occurrence ratios and comparing rules within partitions, the system identifies and eliminates redundant rules while preserving those that contribute to prediction accuracy, thus reducing rule set size without sacrificing reliability.
Solution Approach 2:
The patent discards redundant rules based on co-occurrence analysis and recovers only the essential rules needed for accurate predictions. This selective discarding process reduces rule set complexity while maintaining the predictive power of the system by keeping rules that provide unique valuable information.
3Reliability
If comprehensive rule sets are maintained, then all patterns are covered, but interpretation complexity increases and transparency decreases
Solution Approach 1:
The patent extracts only the essential, non-redundant rules from the comprehensive rule set through pruning. By removing rules that do not add unique value, the system maintains coverage of all important patterns while significantly improving interpretability, as fewer rules are easier for humans to understand and inspect.
4Device complexity
If rule pruning is performed aggressively, then rule set size is reduced, but prediction accuracy may deteriorate
Solution Approach 1:
The patent incorporates feedback mechanisms where pruning decisions are based on calculated co-occurrence ratios and partition comparisons. This feedback ensures that only truly redundant rules are removed, while rules that contribute to prediction accuracy are preserved. The iterative nature of the pruning process allows continuous refinement, maintaining accuracy while reducing rule set size.
Data Source
AI summary
System receives inputs, each input associated with a label and having features, creates a rule for each feature, each rule including a feature and a label, each rule stored in a hierarchy, and distributes each rule into a partition associated with a label or another partition associated with another label. System identifies a number of inputs that include a feature for a rule in the rule partition, and identifies another number of inputs that include both the feature for the rule and another feature for another rule in the rule partition. System deletes the rule from the hierarchy if the ratio of the other number of inputs to the number of inputs satisfies a threshold and an additional number of inputs that includes the other antecedent feature is at least as much as the number. System predicts a label for an input including features by applying each remaining rule to the input.