Duplicate Predictions

How is the duplicate badge triggered.

Some competitions have a duplicate detection feature that flags all models that have a prediction correlation above and makes them ineligible for prizes.

The correlation is computed using pandas.DataFrame.corr(method="spearman").

Grouping

The prediction are first grouped between:

• user, to avoid duplication between multiple models

• team, to avoid duplicate between all models of every team member

The correlation function is then called for each prediction against each other.

Keeping

Predictions are then grouped into correlated pairs to isolate them.

The model in a pair is always kept. Other models are treated as copies.

Example

• If A & B are considered duplicates, and C & D are also considered duplicates, then A and C will be retained and B and D will have the duplicate badge.

• However, if B & C are also considered duplicates, then only A will be retained. B, C, D will have the duplicate badge.

• In some cases, even if A and D are not considered to be duplicates, because of the link from A to B to C to D, the duplicate badge is still displayed.