CatBoost: Ordered Target Encoding
Intermediate
The problem, from first principles
Target encoding replaces a category with the average label for that category. Done naively on the training set, each sample's own label leaks into its feature and the model overfits. CatBoost (Prokhorenkova et al., 2018) avoids this by encoding each sample using only the samples that came before it.
From theory to code
Implement ordered_target_encoding(cats, ys, prior, a), which encodes each sample from earlier samples in its category, blended with the prior.
Constraints
- Process the samples in the given order.
Hints
Hint 1
Keep a running sum and count per category. Encode the current sample first, then update the sums with its label.
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.