Imbalanced Dataset Defender

Current mission

Fraud Shield

Catch rare fraudulent transactions while keeping false alarms under control.

Target recall80%

Max false-positive rate25%

Maximum cost120

Time remainingUnlimited

Score0

mission points

Level1

defender rank

Budget100

resource credits

Trials0 / 8

training attempts

Best F10.000

saved locally

Streak0

successful missions

Dataset Defence Field

Pan: drag · Zoom: wheel · Reset: R

Majority Minority Synthetic

Live Performance

Metrics Comparison

Baseline Confusion Matrix

untreated training

Defended Confusion Matrix

current strategy

Method Comparison Laboratory

Method	Recall	Precision	F1	Cost	Samples

Dataset Generator

build each scenario

Samples 500

Minority rate 10%

Class overlap 35%

Label noise 3%

Difficulty

Random seed

Balancing Strategy

Cost: 0 credits

Defence method

Target ratio 50%

Strength 70%

Neighbours 5

Minority weight 3.0×

False-negative cost 10

False-positive cost 2

Threshold 0.50

Augment noise 8%

Balance training data only

Preserve untouched test distribution

Model Training

Ready

Model

Training split 70%

Epochs 80

Learning rate 0.08

Five-fold cross-validation estimate

Defender Guidance

Generate a dataset, inspect the imbalance, and train a baseline before applying a defence strategy.

Mission completion0%

Achievements

Round Log

Learning Guide

Why accuracy misleads

A model predicting only the majority class can look accurate while missing every rare event.

Keep testing honest

Resample only training data. An altered test set hides real-world class imbalance and causes leakage.

Choose suitable metrics

Prioritise minority recall, F1, balanced accuracy, PR-AUC, and business error costs.