Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Minimising changes to audit when updating decision trees

Anj Simmons, Scott Barnett, Anupam Chaudhuri, Sankhya Singh, Shangeetha Sivasothy

TL;DR

Updates to interpretable decision trees on new data are performed while minimizing auditing changes. The Keep-Regrow algorithm greedily chooses at each split between keeping the existing condition or regrowing that subtree, optimizing the joint loss $L(r_{t-1}, r_t, D_t) = f(r_t, D_t) + \alpha c(r_t) + \beta \Delta(r_{t-1}, r_t)$. Pruning is integrated after regrowth with Preserve Split and Terminate options to control tree complexity and auditability. Evaluations on six UCI datasets show Keep-Regrow achieves a favorable trade-off between accuracy, tree size, and audit similarity, with practical defaults $\alpha=5$ and $\beta=1$. The work positions itself among incremental decision trees and interpretability research, and outlines future work in human-in-the-loop testing and lifecycle integration.

Abstract

Interpretable models are important, but what happens when the model is updated on new training data? We propose an algorithm for updating a decision tree while minimising the number of changes to the tree that a human would need to audit. We achieve this via a greedy approach that incorporates the number of changes to the tree as part of the objective function. We compare our algorithm to existing methods and show that it sits in a sweet spot between final accuracy and number of changes to audit.

Minimising changes to audit when updating decision trees

TL;DR

Updates to interpretable decision trees on new data are performed while minimizing auditing changes. The Keep-Regrow algorithm greedily chooses at each split between keeping the existing condition or regrowing that subtree, optimizing the joint loss . Pruning is integrated after regrowth with Preserve Split and Terminate options to control tree complexity and auditability. Evaluations on six UCI datasets show Keep-Regrow achieves a favorable trade-off between accuracy, tree size, and audit similarity, with practical defaults and . The work positions itself among incremental decision trees and interpretability research, and outlines future work in human-in-the-loop testing and lifecycle integration.

Abstract

Interpretable models are important, but what happens when the model is updated on new training data? We propose an algorithm for updating a decision tree while minimising the number of changes to the tree that a human would need to audit. We achieve this via a greedy approach that incorporates the number of changes to the tree as part of the objective function. We compare our algorithm to existing methods and show that it sits in a sweet spot between final accuracy and number of changes to audit.
Paper Structure (20 sections, 10 equations, 1 figure, 1 table)

This paper contains 20 sections, 10 equations, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Objective Function
  3. Keep-Regrow Algorithm
  4. Keep
  5. Regrow
  6. Prune
  7. Preserve Split
  8. Terminate
  9. Example
  10. Datasets
  11. Evaluation
  12. Exploration of Alpha Parameter
  13. Exploration of Beta Parameter
  14. Exploration of Algorithm
  15. Exploration of behaviour over time
  16. ...and 5 more sections

Figures (1)

  • Figure 1: Initial decision tree at t=0 (left), and updated tree at t=1 (right) produced by our algorithm. Note how our algorithm 'keeps' Node 1 and Node 2 to minimise changes to audit and 'regrows' Node 3 for improved accuracy.