Relevance-Aware Rule: Structural Deletion of Irrelevant Conditions in Decision Trees
Merged summary
TL;DR — A method for reliably removing irrelevant conditions (IRCs) from decision-tree if–then rules by exploiting the structural mechanism of binary splitting, simplifying rules without hurting reliability. It matters because it improves interpretability of tree-based models while preserving predictive fidelity.
- Identifies that IRCs stem from tree-splitting structure and persist even in modern optimal sparse tree algorithms, which prior deletion methods ignore.
- Establishes theory showing a binary split shifts class proportions oppositely across siblings, creating "C1-links" and "C0-links."
- Relative to each leaf, mismatched links (those increasing the opposite class's proportion) are flagged as suspicious IRC candidates rather than deleted outright.
- Candidates are diagnosed via prediction-reliability assessment; only structurally and empirically irrelevant conditions are removed, yielding substantial simplification with reliability preserved (per reported experiments).
Sources (1)
Relevance-Aware Rule: Structural Deletion of Irrelevant Conditions in Decision Trees
TL;DR — A method for reliably removing irrelevant conditions (IRCs) from decision-tree if–then rules by exploiting the structural mechanism of binary splitting, simplifying rules without hurting reliability. It matters because it improves interpretability of tree-based models while preserving predictive fidelity.
- Identifies that IRCs stem from tree-splitting structure and persist even in modern optimal sparse tree algorithms, which prior deletion methods ignore.
- Establishes theory showing a binary split shifts class proportions oppositely across siblings, creating "C1-links" and "C0-links."
- Relative to each leaf, mismatched links (those increasing the opposite class's proportion) are flagged as suspicious IRC candidates rather than deleted outright.
- Candidates are diagnosed via prediction-reliability assessment; only structurally and empirically irrelevant conditions are removed, yielding substantial simplification with reliability preserved (per reported experiments).