Converting CatBoost model into yet another CatBoost model without specified float features
A CatBoost model based on oblivious trees can be modified in way to remove float features by replacing them with constants. There are two types of replacing implemented:
- Carry: replace features subset by fixed values. Resulting model produce n-dimensional scores (where n is number of values specified for each feature).
- Uplift: replace features subset by value pair. Resulting model produce one dimensional score equal to score difference between carrying model by second and first value from pair.
The functions for carry and uplift are defined in the catboost/libs/carry/carry.h. These functions produce CatBoost models wich can be applyed as usual.
NOTE: Only float features not used in any CTR are supported
Carry
Given a model model and a grid factorValues of feature values, CarryModelByFeatureIndex builds a model which takes samples without features factorFeatureIndexes, and returns a vector of predictions of model with features factorFeatureIndexes set to values factorValues.
You may also specify the features for carry by flat index, name, and position using other carry functions.
NOTE: Model size increases proportional to row count in grid
factorValues
TFullModel CarryModelByFeatureIndex(const TFullModel& model, const TVector<int>& factorFeatureIndexes, const TVector<TVector<double>>& factorsValues);
TFullModel CarryModelByFlatIndex(const TFullModel& model, const TVector<int>& factorFlatIndexes, const TVector<TVector<double>>& factorsValues);
TFullModel CarryModelByName(const TFullModel& model, const TVector<TString>& factorNames, const TVector<TVector<double>>& factorsValues);
TFullModel CarryModel(const TFullModel& model, const TVector<TFeaturePosition>& factors, const TVector<TVector<double>>& factorValues);
Uplift
Given a model model and two features samples baseValues and nextValues of values, UpliftModelByFeatureIndex builds a model which takes samples without features factors, and returns the difference of predictions of model with features factors set to baseValues and nextValues.
You may also specify the features for uplift by flat index, name, and position using other uplift functions.
TFullModel UpliftModelByFeatureIndex(const TFullModel& model, const TVector<int>& factors, const TVector<double>& baseValues, const TVector<double>& nextValues);
TFullModel UpliftModelByFlatIndex(const TFullModel& model, const TVector<int>& factors, const TVector<double>& baseValues, const TVector<double>& nextValues);
TFullModel UpliftModelByName(const TFullModel& model, const TVector<TString>& factors, const TVector<double>& baseValues, const TVector<double>& nextValues);
TFullModel UpliftModel(const TFullModel& model, const TVector<TFeaturePosition>& factors, const TVector<double>& baseValues, const TVector<double>& nextValues);