AdaBoost Weight Calculator

Calculation Results

Results appear here after calculation.

Not calculated

Weighted error—

Weighted accuracy—

Learner weight (alpha)—

Normalization constant—

Highest sample weight—

Lowest sample weight—

Effective sample size—

Total normalized weight—

Enter sample data, then select Calculate.

#	Sample	Actual	Predicted	Status	Current weight	Multiplier	Raw new weight	Normalized weight	Change	Change %
No calculated rows.

Weight Distribution

Detailed Output

No results yet.

Calculator Settings

Weak learner name

Calculation mode

Weighted error source

Manual weighted error

Learning rate (η)

Number of classes (K)

Boosting round

Error clipping epsilon

Maximum absolute alpha

Decimal precision

Weight update rule

Initial weight method

Number display

Normalize initial weights

Normalize updated weights

Clip error for stability

Training Samples

Add actual labels, predictions, and current sample weights.

#	Sample name	Actual label	Predicted label	Current weight	Remove

Import CSV file

Use four columns: sample name, actual label, predicted label, and weight.

Paste spreadsheet data

Accepted columns: sample name, actual label, predicted label, weight. Commas or tabs are supported.

Final Ensemble Prediction

Enter each weak learner prediction and alpha. Binary predictions usually use -1 and 1.

Prediction mode

Positive label

Negative label

#	Learner name	Prediction	Alpha	Weighted vote	Remove

No ensemble prediction yet.

Boosting-Round History

Commit each calculated round, then reuse updated weights for the next round.

Round	Learner	Mode	Error	Accuracy	Alpha	Z	Min weight	Max weight	Average weight	ESS	Misclassified weight	Misclassified share	Cumulative alpha
No committed rounds.

Formula Used

Weighted error: εₜ = Σ wᵢ I(yᵢ ≠ hₜ(xᵢ)) / Σ wᵢ

Binary learner weight: αₜ = η × ½ ln((1 − εₜ) / εₜ)

SAMME learner weight: αₜ = η × [ln((1 − εₜ) / εₜ) + ln(K − 1)]

Updated weight: wᵢ(new) = wᵢ × exp(−αₜ) when correct, or wᵢ × exp(αₜ) when incorrect

Normalization: Zₜ = Σ wᵢ(new), then wᵢ(normalized) = wᵢ(new) / Zₜ

Effective sample size: ESS = 1 / Σ wᵢ²

How to Use

Select binary AdaBoost or multiclass SAMME.
Enter actual labels, predictions, and current weights.
Choose automatic or manual weighted error.
Set learning rate, clipping, precision, and normalization.
Select Calculate to update every sample weight.
Commit a round to build a boosting history.
Use the ensemble section for final weighted predictions.

Worked Example

Sample	Actual	Predicted	Initial weight	Classification
A	1	1	0.20	Correct
B	1	-1	0.20	Incorrect
C	-1	-1	0.20	Correct
D	-1	-1	0.20	Correct
E	1	1	0.20	Correct

The weighted error is 0.20. Binary alpha is approximately 0.693147. The misclassified sample gains greater relative importance.

Frequently Asked Questions

What does alpha represent?

Alpha measures a weak learner’s influence. Larger positive values create stronger ensemble votes.

Why are misclassified samples increased?

AdaBoost focuses later learners on difficult observations. Their relative weights therefore become larger.

What happens when error equals 0.5?

A binary learner performs like random guessing. Its calculated influence approaches zero.

Can alpha become negative?

Yes. This occurs when binary weighted error exceeds one-half.

Why clip weighted error?

Clipping prevents logarithms of zero. It also limits extreme numerical values.

What is SAMME?

SAMME extends discrete AdaBoost to multiclass classification. It includes the class-count correction term.

Must weights sum to one?

Normalized weights usually sum to one. The weighted error ratio still works with positive unnormalized weights.

What does effective sample size show?

ESS estimates how concentrated the weight distribution is. Smaller values indicate stronger concentration.

Can I simulate several rounds?

Yes. Commit a round, change predictions, and reuse updated weights.