Incognition logo
Home
Learn
Practice
Leaderboard
0
VisualizationsApplied Machine Learning

Global Surrogate Tree

Pick a tree depth and watch a small decision tree copy a random forest's loan decisions. Ringed dots mark where the copy and the forest disagree.

A small tree copying the forest
Each dot is an applicant, coloured by the forest's answer. The tree learns from these colours, not from the true outcomes.
forest approvesforest refusesthe copy disagrees
55060065070075080020406080100Credit Score →Income ($k) →700$50k$90k760
Tree depth
Rules: 2 (at most 2)Matches the forest: 19/25 = 76%
The copy's rulesmatches
  • IF Credit > 700 THEN Approve8/11
  • IF Credit ≤ 700 THEN Refuse11/14
  • IF Credit > 700 AND Income > $50k THEN Approve7/7
  • IF Credit > 700 AND Income ≤ $50k THEN Refuse3/4
  • IF Credit ≤ 700 AND Income > $90k THEN Approve2/2
  • IF Credit ≤ 700 AND Income ≤ $90k THEN Refuse11/12
  • IF Credit > 700 AND Income > $50k THEN Approve7/7
  • IF Credit > 760 AND Income ≤ $50k THEN Approve1/1
  • IF 700 < Credit ≤ 760 AND Income ≤ $50k THEN Refuse3/3
  • IF Credit ≤ 700 AND Income > $90k THEN Approve2/2
  • IF Credit ≤ 700 AND Income ≤ $90k THEN Refuse11/12

One question copies the forest on 19 of 25 applicants. The 6 ringed dots are where its rules tell the wrong story about the forest.

Two levels, 4 rules, and only 2 ringed dots. The first rule is the same as Priya's cheat sheet: Credit > 700 AND Income > $50k.

A third level fixes 1 more dot but adds a 5th rule. The last ringed dot sits among refusals, so one more question can't pull it out.

Customize