Game settings
Drag each column card into a role zone. Keyboard users can select a card, then press 1–6 to assign its role.
Dataset preview
Use values, data types, missingness, uniqueness, and prediction timing to decide each role.
| Column | Type | Example values | Missing | Unique | Selected role |
|---|
Performance dashboard
Custom challenge builder
Enter one column per line using: name | type | samples | role | explanation. Valid roles: feature, label, identifier, metadata, leakage, irrelevant.
Learning guide
Feature
An input available when a prediction is made. It should provide reusable information about the outcome.
Target label
The outcome a supervised model learns to predict. A challenge normally has one primary target.
Identifier
A row, customer, transaction, or device key. It often memorizes records instead of learning patterns.
Metadata
Context used for auditing, grouping, or tracing data. It may not belong in the predictive model.
Leakage
Information unavailable at prediction time or derived from the outcome. Leakage creates unrealistically strong evaluation results.
Irrelevant
A column unrelated to the objective or too noisy to justify inclusion. Removing it can simplify the model.
Regression
Predicts a continuous numeric value, such as price, demand, duration, or temperature.
Classification
Predicts a category, such as churned or retained, approved or declined, or one of several classes.
Unsupervised learning
Finds structure without a target label. Clustering challenges may correctly contain no label.
Instructions and frequently asked questions
How do I play?
Read the prediction objective. Inspect each column card and dataset sample. Drag every card into the best role zone, then submit your classification.
Can a column have different roles?
Yes. A timestamp can be a feature, metadata, identifier, or leakage depending on the objective and prediction moment.
Why are identifiers usually excluded?
Identifiers are often unique and do not generalize. They can encourage memorization, unstable patterns, and privacy risks.
What makes a column leakage?
A leaking column contains future information, post-outcome information, or a transformation that directly reveals the target.
Can sensitive attributes be features?
Sometimes they are retained for fairness auditing but excluded from training. Their treatment depends on law, policy, ethics, and the use case.
What happens in clustering?
Clustering is unsupervised. There may be no target label, so selecting a label can be an error.
How is the score calculated?
Correct roles earn points. Fast answers, streaks, leakage detection, and no-hint rounds add bonuses. Incorrect roles and hints reduce points.
Does progress persist?
Yes. Scores, charts, settings, achievements, and best streaks are stored locally in your browser.