LoginSign Up
← Back to all sparks
E

e2tree

ANALYTICS
Velocity0.0

R package e2tree by massimoaria — release notes from GitHub.

github.com/massimoaria/e2tree ↗

The explainable-ensemble-tree package now measures whether its own explanations are faithful.

explainable-aiensemble-methodsdecision-treesr-packagemodel-interpretabilitycatboost
◆Current state
e2tree builds a single interpretable tree that approximates a fitted ensemble, working from the proximity structure the ensemble induces between observations. The 1.0.0 release added the piece that had been missing: a Goodness of Interpretability index quantifying how well the approximating tree reconstructs the ensemble's own proximity matrix, with a permutation test for significance. Interactive visualisation and a C++ backend with OpenMP parallelism arrived alongside, and support now spans ranger and CatBoost as well as the original targets.
◆Where it's heading
Development has moved from producing an explanation to defending it. The GoI index and its permutation test change the package's claim from here is a tree that resembles your ensemble to here is how closely it resembles it and whether that could have happened by chance — the question a reviewer asks of any surrogate model. Around that, the work is engineering: the proximity matrix construction moved from R-level parallel loops into C++ with thread-level parallelism, and recent releases have been absorbing the awkwardness of supporting multiple ensemble backends, where a multi-class CatBoost objective returns a score matrix where a vector was expected. Interactive visNetwork output and standalone HTML export point at explanations meant to be shared rather than only inspected.
◆Prediction
Given how much recent effort has gone into per-backend adapters, expect further work on ensemble compatibility; the entries do not indicate whether the interpretability index is heading toward comparing surrogate trees against each other.

◆Recent moves

  1. 3mo ago

    CatBoost multi-class and loss-function handling repaired

    Two hard failures in the CatBoost adapter get fixed: multi-class objectives return a per-class score matrix that broke leaf-proxy assignment outright, and some CatBoost releases return the loss function as a list rather than a string, which crashed detection. Neither is glamorous, but supporting several ensemble backends means inheriting each one's return-type quirks, and both bugs made the adapter unusable rather than merely wrong.

    View source ↗
  2. 4mo ago

    A significance-tested measure of explanation fidelity

    ⚡ SPARK

    The release that gives the method a way to be checked: goi() measures how well the estimated tree reconstructs the original ensemble's proximity matrix, goi_perm() tests that measure against a permutation distribution, and goi_analysis() returns both. Interactive visualisation and a C++ parallel backend ship in the same release, but the index is what changes what this package can claim.

    View source ↗
  3. 1y ago

    ranger models supported

    Support for ranger widened the set of ensembles the method could be applied to, the start of the multi-backend surface whose per-library quirks later releases have been absorbing. Plot improvements accompany it.

    View source ↗