roc-best-threshold — Reddit case

Reddit-inspired moderation model produces continuous risk scores while the operations team needs a binary decision threshold. You need to evaluate candidate thresholds and select the one that maximizes Youden’s J statistic.

Choose the decision threshold that maximises Youden's $J=\text{TPR}-\text{FPR}J=\backslash text\{TPR\}-\backslash text\{FPR\}$. Candidate thresholds are the distinct scores; an example is predicted positive when score >= threshold. Labels y are 0/1, and if several thresholds give the same $JJ$ the smallest one is returned.

Implement solve(y,scores).

Returns. Return the chosen threshold (one of the scores) as a float. A rate whose denominator is zero (no positives or no negatives) counts as 0.

Implement solve(y, scores).

Examples

Example 1

Input:

solve([1, 1, 0, 0], [0.9, 0.7, 0.6, 0.2])

Output:

0.7

Example 2

Input:

solve([0, 0, 1, 1], [0.1, 0.4, 0.35, 0.8])

Output:

0.35