Interpretability: what is measured, and what is only searched for¶
This page states the project's position on interpretability. It applies to anything written about results from this framework.
Reported measures¶
Interpretability is reported with established structural proxies only [Freitas 2014; Piltaver et al. 2016]:
| Measure | Definition | Where it comes from |
|---|---|---|
| Node count | Internal nodes + leaves | ga_tree_complexity, tree_.node_count (CART) |
| Leaf count | Number of rules the tree encodes | same |
| Mean decision-path length | Depth of the leaf each sample lands in, averaged over the samples | same — weighted by the data, not by leaf count |
| Distinct features used | Features that appear in at least one split | same |
These are properties of the tree that anyone can recompute. None of them is claimed to be interpretability; they are the size- and length-based proxies that the comprehensibility literature most often uses, and every one is reported for every method, including CART.
The frontier benchmark uses node count as its complexity axis, fixed before the benchmark was run.
The composite score is a search heuristic¶
FitnessCalculator combines four terms into one "interpretability" number used by the
weighted-sum fitness:
| Term | What it computes | Why it is not an outcome measure |
|---|---|---|
node_complexity |
Decreasing function of node count | Redundant with node count, which is reported directly |
feature_coherence |
1 − n_used / n_total |
Scales with dataset dimensionality: a 30-feature dataset gets a high score for free |
tree_balance |
Balance of left/right subtree depths | No evidence that balance helps a human read a tree |
semantic_coherence |
Standard deviation of the depths at which features appear | Invented for this project; no grounding in the literature and no human study |
The composite steers the single-objective search toward small trees, and in that role it works. It has no claimed validity as a measure of human interpretability, and it appears in no reported result. Validating any of these terms would need a forward-simulation user study [Doshi-Velez & Kim 2017; Lipton 2018].
A consequence worth knowing¶
The weighting in the composite is the dominant factor in how big a single-objective GA tree
ends up. At configs/paper.yaml's weights a tree must gain about 8 accuracy points to justify
growing from a stump to the size of a tuned CART tree (22.6 points at configs/fast.yaml).
So a single GA tree and a single CART tree compared on accuracy alone are usually two points
at opposite ends of the size axis. Compare frontiers, or check the leaf column, before reading
an accuracy gap.