CatBoost: The Target Statistic Formula
Beginner
The problem, from first principles
A category seen only once or twice gives an unreliable mean. Blending that mean with a prior, weighted by a, keeps the encoding stable when data is sparse.
From theory to code
Implement target_statistic(sum_y, count, prior, a), returning (sum_y + a * prior) / (count + a).
Constraints
- With no history the statistic equals the prior.
Hints
Hint 1
Add the prior's pseudo-count to the label sum and the denominator.
Keywords
research-papersclassical-mlboostingcatboost
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.