[ Tech stack ]
CatBoost
Yandex's gradient-boosting library that handles categorical features with no preprocessing.
CatBoost trains symmetric (oblivious) decision trees with ordered boosting: each example's residuals come from a model that never saw it, which corrects the prediction shift of classic boosters. It encodes categorical features itself with ordered target statistics and automatic feature combinations, handles text and embedding features, and performs well with its default hyperparameters. Available for Python, R, Java and C++ with GPU acceleration, it's our choice for data with many or high-cardinality categories, when we want a robust model without a long tuning phase.
[ Why CatBoost at Dexon ]
What this technology does well, and why we use it.
Typical usage: Customer scoring, insurance pricing, recommendation and classification on category-heavy data.
- 01
Ordered boosting: less overfitting on small and medium-sized datasets.
- 02
Native categorical encoding with ordered target statistics and automatic combinations.
- 03
Symmetric trees: very fast inference, compact model, robust out of the box.
- 04
Text and embedding features, multi-GPU training, ONNX, CoreML and C++ export.
[ Complementary technologies ]
The building blocks we often mobilise alongside
A stack rarely exists alone. Here are the technologies Dexon most often pairs with this one, through pipeline habits, usage similarity or internal mastery. Click on a building block to see its scope.
[ Reassurance ]
- 0+
- custom projects delivered
- 30
- engineers, designers, project managers
- 80 %
- from top French schools
- 24 h
- average reply time
[ They trust us ]
More than 100 French and European companies trust us
[ Press ]
They talk about us
Discover what the business and tech press say about us. An outside reading of our work and our innovations.
