Challenge dashboard
Find the split that reduces uncertainty most.
Score
0
Streak
0
Round
1/10
Best score
0
Session progress10%
Challenge settings
Dataset
Inspect features and class labels.
Split workspace
Keyboard: N new, T test, H hint.
Feedback
Choose a feature and split value, then test the split.
Information Gain = Parent Entropy − Weighted Child EntropyParent impurity
—
Weighted child
—
Gain
—
Split info
—
Animated split canvas
Records move from the parent into child nodes.
Candidate comparison
Plotly analytics
Learning guide
1. Measure the parent.
Calculate entropy or Gini impurity before splitting.
Calculate entropy or Gini impurity before splitting.
2. Divide records.
Apply one feature condition to create left and right groups.
Apply one feature condition to create left and right groups.
3. Weight child impurity.
Multiply each child impurity by its record proportion.
Multiply each child impurity by its record proportion.
4. Compare reduction.
The strongest split removes the most uncertainty.
The strongest split removes the most uncertainty.
Entropy(S) = −Σ pᵢ log₂(pᵢ)
Gini(S) = 1 − Σ pᵢ²
Gain = I(parent) − Σ (|child| / |parent|) × I(child)Achievements
Results and data tools
Generated: 2026-08-10T12:58:20+00:00
Frequently asked questions
Why is the highest information gain preferred?
It creates the largest reduction in class uncertainty. The resulting child nodes are usually purer.
Can information gain favour many-valued features?
Yes. Gain ratio can reduce that preference by dividing gain by split information.
How are numerical thresholds generated?
The game tests midpoints between sorted unique values. Invalid small leaves are removed.
What happens when gains are tied?
The game accepts values within a small tolerance. Simpler balanced splits receive explanatory preference.
Does the game support multiple classes?
Advanced datasets may contain three classes. Entropy and Gini work for every class count.
Where is progress stored?
Scores, achievements, and history use local browser storage. No account is required.