Skip to content
AI360Xpert
Beta
Core ML

Visual explainer

Ensemble Stacking

A tree says 5.2 and a linear model says 3.1, so a trained meta-model weights them 0.1 and 0.9 to land at 3.31 near the truth of 3.5.

A fixed average lands at 4.15 while the truth sits at 3.5.
A fixed average lands at 4.15 while the truth sits at 3.5.

A tree shouts 5.2 while a linear model whispers 3.1, and the fixed average lands at 4.15 against a truth of 3.5. Equal weights cannot hear which model owns this street this winter, every winter.

Honest Evidence Only

Each model predicts the one fold it never trained on.
Each model predicts the one fold it never trained on.

You split streets into four folds and make each model predict the fold it never trained on. The judge sees real mistakes, never memorized answers, which is the entire trick.

Predictions Become Rows

Weights 0.1 and 0.9 blend 5.2 and 3.1 into 3.31.
Weights 0.1 and 0.9 blend 5.2 and 3.1 into 3.31.

Each street becomes one row: tree 5.2, linear 3.1, truth 3.5. The meta-model learns weights 0.1 and 0.9, blending 0.52 plus 2.79 into 3.31. That 3.31 sits 0.19 from truth; the flat average sat 0.65 away.

Trust Shifts by Street

Freezing streets trust the tree at 0.8 while mild streets trust the line at 0.85.
Freezing streets trust the tree at 0.8 while mild streets trust the line at 0.85.

On freezing streets the tree earns weight 0.8; on mild ones the linear model earns 0.85. The blend follows the weather instead of a fixed recipe, and that adaptivity is the whole point.

Error Nearly Vanishes

Error falls from 0.65 to 0.19 on the same eight streets.
Error falls from 0.65 to 0.19 on the same eight streets.

The fixed average misses by 0.65 across the same eight streets. The stack misses by 0.19, because it learned who to trust where.

Keep the Judge Simple

Two weights beat two hundred until streets number in the thousands.
Two weights beat two hundred until streets number in the thousands.

A linear meta-model has two weights and nothing to memorize, so default to it. When you have thousands of streets, you may audition something fancier with cross-validation.

Where It Breaks

Three identical trees vote as one while charging for three.
Three identical trees vote as one while charging for three.

Three identical trees make the same mistakes on the same streets, so the meta-model learns nothing. You pay triple compute for one opinion stated thrice.

The Quick Version

  • Fixed averages cannot hear expertise.
  • Judges train on unseen folds only.
  • Weights 0.1 and 0.9 blend to 3.31.
  • Cold streets trust trees; mild trust lines.
  • Error falls 0.65 to 0.19.
  • Default linear judges; fancy needs data.
  • Identical models teach nothing, cost triple.