Attributes¶
feature_importances_¶
Purpose¶
Return the importance of each input feature, in the order of the input columns. The values sum to 1.
Each deployed effect's share of the model's variance (its Sobol index under the reference
measure) is split equally among the features it involves, as in the Shapley split of
predict_contributions. A multiclassification model averages this over its classes. It is read
from the model and needs no data. See Feature importance.
Type¶
numpy.ndarray
required_columns¶
Purpose¶
The input columns needed to apply the model, in the order of the training data. Prediction methods read these columns by name, and extra columns are ignored. The exposure column is not among them. Use it to select the columns of a LazyFrame before collecting it:
Type¶
list of strings
n_features_in_¶
Purpose¶
The number of features seen during fit.
Type¶
int
feature_names_in_¶
Purpose¶
The names of the features seen during fit. Only set when X had column names (for example, a
polars DataFrame): check with getattr(model, "feature_names_in_", None).
Type¶
numpy.ndarray
categories_¶
Purpose¶
The levels of each categorical feature seen during fit, keyed by feature name. The list
includes the levels pooled into the "<rare>" level, and None for the missing level when the
fit saw missing values. The values are the strings t-boost matches on (a numeric category 1 is
"1").
Type¶
dict
link¶
Purpose¶
The link between the raw score and the prediction: identity, log, logit or, for a
multiclassification model, softmax.
Type¶
string
n_trees_¶
Purpose¶
The largest number of trees kept by a bag. This number can differ from the value specified in
the n_trees training parameter in the following cases:
- The training is stopped by early stopping.
- No split clears
min_split_gain. - A callback stops the training.
None for a multiclassification model.
Type¶
int
n_trees_per_bag_¶
Purpose¶
The number of trees each bag kept. None for a multiclassification model.
Type¶
list of ints
stopping_reason_¶
Purpose¶
Why the training stopped, summarized over the bags: "callback" if a callback stopped it, else
"early_stopping" if a bag was stopped by early stopping, else "no_split" if a bag
ran out of splits, else "max_trees". None for a multiclassification model.
Type¶
string
stopping_reason_per_bag_¶
Purpose¶
Why each bag stopped: "early_stopping", "max_trees", "no_split" or "callback". None for
a multiclassification model.
Type¶
list of strings
evals_result_¶
Purpose¶
Return the values of metrics calculated during the training: the mean deviance of every
iteration of every bag, when an eval_set or callbacks were passed to fit.
Output format:
{"train": {"deviance": [[value_1, value_2, ...], ...]}, "eval": {"deviance": [[value_1, value_2, ...], ...]}}
with one list per bag. The "eval" curve (the validation objects) costs nothing extra. The
"train" curve takes an extra pass per iteration, so it is calculated only when callbacks are
given. Without either, the dictionary is empty.
Type¶
dict
pruning_report_¶
Purpose¶
The record of the table selection: the selector used (selector), the candidate tables and
their scores (table_scores), the kept tables (kept), the scored path (path) and, when they
ran, the reports of banding (banding) and of the budgets (table_budget, box_budget). Only
set when prune=True.
Type¶
dict
graduation_report_¶
Purpose¶
The graduation diagnostics, one entry per table: the features, the smoothing strength (alpha)
and whether the smoothing was rejected. Only set when graduation is turned on.
Type¶
list of dicts
binding_report_¶
Purpose¶
Which of the parameters set away from their defaults took effect on the fit. See check_bindings.
Type¶
list of dicts
delta_step_gate_¶
Purpose¶
What the guard of max_delta_step_gated
saw: whether it engaged (engaged, engaged_round, bags_engaged), the smallest log ratio of a
predicted rate to the mean rate (min_log_rate_ratio) and the guard's settings. None when no
guard was armed.
Type¶
dict
metadata¶
Purpose¶
Your own JSON metadata, saved with the model by to_bytes and to_json and returned unchanged by
from_bytes and from_json. t-boost never reads it: it has no effect on the training, the
predictions or the tables, and a new fit keeps it. The keys must be strings and the values
JSON-serializable (no NaN or infinity), otherwise saving the model raises SerializationError.
Type¶
dict