Train and store a classification model to be loaded at a later point for prediction.
The output will consist of a new column with the trained model’s predictions on the training data,
as well as a saved and named model file that can be used in other projects for prediction of new data.
Optionally, if a second output column name is provided, the model’s predicted probabilities will also be
returned.
A detailed guide on how to configure this step for model tuning and performance evaluation can be found
here.
The following examples show how the step can be used in a recipe.
Train a classification model with default parameters. By default, a Catboost model will be trained, but this can be changed to any of the supported models by specifying the model parameter (see below for details):
The following are the inputs expected by the step and the outputs it produces. These are generally
columns (ds.first_name), datasets (ds or ds[["first_name", "last_name"]]) or models (referenced
by name e.g. "churn-clf").
One or two columns containing the model’s predictions. If two column names are provided, the second column
will contain the model’s predicted probabilities.
The following parameters can be used to configure the behaviour of the step by including them in
a json object as the last “input” to the step, i.e. step(..., {"param": "value", ...}) -> (output).
Name of the positive class.
In binary classification, usually the class you’re most interested in, for example the label/class
corresponding to successful lead conversion in a lead score model, the class corresponding to a
customer who has churned in a churn prediction model, etc.
If provided, will automaticall measure the performance (accuracy, precision, recall) of the model on this
class, in addition to averages across all classes. If not provided, only summary metrics will be reported.
Maximum number of classes in the target variable.
If there are more classes than this, the least frequent classes will be grouped together into a single class
called “others”. Reducing the number of classes in the target variable can help improve model performance,
especially when the number of classes is very large, some classes are very rare, or the dataset doesn’t have
sufficient samples for all classes. Raising this significantly might lead to much longer training times.
Importance of each feature in the model.
Whether and how to measure each feature’s contribution to the model’s predictions. The higher the value,
the more important the feature was in the model. Only relative values are meaningful, i.e. the importance
of a feature relative to other features in the model.
Also note that feature importance is usually meaningful only for models that fit the data well.
The default (null, true or "native") uses the classifier’s native feature importance measure, e.g.
prediction-value-change in the case
of Catboost, Gini importance
in the case of scikit-learn’s DecisionTreeClassifier, and the mean of absolute coefficients in the case of
logistic regression.
When set to "permutation", uses permutation importance,
i.e. measures the decrease in model score when a single feature’s values are randomly shuffled. This is
considerably slower than native feature importance (the model needs to be evaluated an additional k*n times,
where k is the number of features and n the number of repetitions to average over). On the positive side it is
model-agnostic and doesn’t suffer from bias towards high cardinality features (like some tree-based feature
importances). On the negative side, it can be sensitive to strongly correlated features, as the unshuffled
correlated variable is still available to the model when shuffling the original variable.
When set to false, no feature importance will be calculated.
Toggle encoding of feature columns.
When enabled, Graphext will auto-convert any column types to the numeric type before
fitting the model. How this conversion is done can be configured using the feature_encoder
option below.
If disabled, any model trained in this step will assume that input data
is already in an appropriate format (e.g. numerical and not containing any missing values).
Configures encoding of feature columns.
By default (null), Graphext chooses automatically how to convert any column types the model
may not understand natively to a numeric type.
A configuration object can be passed instead to overwrite specific parameter values with respect
to their default values.
Category encoder.
May contain either a single configuration for all categorical variables, or two different configurations
for low- and high-cardinality variables. For further details pick one of the two options below.
Maximum number of unique categories to encode.
Only the N-1 most common categories will be encoded, and the rest will be grouped into a single
“Others” category.
Multilabel encoder.
Configures encoding of multivalued categorical features (variable length lists of categories,
or the semantic type list[category] for short). May contain either a single configuration for
all multilabel variables, or two different configurations for low- and high-cardinality variables.
For further details pick one of the two options below.
Maximum number of categories/labels to encode.
If a number is provided, the result of the encoding will be reduced to these many dimensions (columns)
using scikit-learn’s truncated SVD.
When applied together with (after a) Tf-Idf encoding, this performs a kind of
latent semantic analysis.
A list of cyclical time features to extract.
“Cycles” are numerical transformations of features that should be represented on a circle. E.g. months,
ranging from 1 to 12, should be arranged such that 12 and 1 are next to each other, rather than on
opposite ends of a linear scale. We represent such cyclical time features on a circle by creating two
columns for each original feature: the sin and cos of the numerical feature after appropriate scaling.
Embedding/vector encoder.
Configures encoding of multivalued numerical features (variable length lists of numbers, i.e. vectors, or the semantic type list[number] for short).
Text encoder.
Configures encoding of text (natural language) features. Currently only allows
Tf-Idf embeddings to represent texts. If you wish
to use other embeddings, e.g. semantic, Word2Vec etc., transform your text column first using
another step, and then use that result instead of the original texts.
Texts are excluded by default from the overall encoding of the dataset. See parameter
include_text_features below to active it.
Parameters to be passed to the text encoder (Tf-Idf parameters only for now).
See scikit-learn’s documentation
for detailed parameters and their explanation.
How many output features to generate.
The resulting Tf-Idf vectors will be reduced to these many dimensions (columns) using scikit-learn’s
truncated SVD.
This performs a kind of latent semantic analysis.
By default we will reduce to 200 components.
Whether to include or ignore text columns during the processing of input data.
Enabling this will convert texts to their TfIdf representation. Each text will be
converted to an N-dimensional vector in which each component measures the relative
“over-representation” of a specific word (or n-gram) relative to its overall
frequency in the whole dataset. This is disabled by default because it will
often be better to convert texts explicitly using a previous step, such as
embed_text or embed_text_with_model.
Number of iterations.
The maximum number of trees that can be built when solving machine learning problems. When using other parameters that limit the number of iterations, the final number of trees may be less than the number specified in this parameter.
Maximum cardinality of variables to be one-hot encoded.
Use one-hot encoding for all categorical features with a number of different values less than or equal to this value. Other variables will be target-encoded. Note that one-hot encoding is faster than the alternatives, so decreasing this value makes it more likely slower methods will be used. See CatBoost details for further information.
The maximum number of features that can be combined when transforming categorical variables.
Each resulting combination consists of one or more categorical features and can optionally contain binary features in the following form: “numeric feature > value”.
The amount of randomness to use for scoring splits.
Use this parameter to avoid overfitting the model. The value multiplies the variance of a random variable (with
zero mean) that is added to the score used to select splits when a tree is grown.
The method for processing missing values in the input dataset.
Possible values:
“Forbidden”:
Missing values are not supported, their presence is interpreted as an error.
“Min”:
Missing values are processed as the minimum value (less than all other values) for the feature.
It is guaranteed that a split that separates missing values from all other values is considered
when selecting trees.
“Max”:
Missing values are processed as the maximum value (greater than all other values) for the feature.
It is guaranteed that a split that separates missing values from all other values is considered when
selecting trees.
Using the Min or Max value of this parameter guarantees that a split between missing values and other
values is considered when selecting a new split in the tree.
Random subspace method.
The percentage of features to use at each split selection, when features are selected over again at random. The value null is equivalent to 1.0 (all features). You can set this to values < 1.0 when the dataset has many features (e.g. > 20) to speed up training.
Configure model validation.
Allows evaluation of model performance via cross-validation
using custom metrics. If not specified, will by default perform 5-fold cross-validation with automatically selected
metrics.
Number of train-test splits to evaluate the model on.
Will split the dataset into training and test set n_splits times, train on the former
and evaluate on the latter using specified or automatically selected metrics.
What proportion of the data to use for testing in each split.
If null or not provided, will use k-fold cross-validation
to split the dataset. E.g. if n_splits is 5, the dataset will be split into 5 equal-sized parts.
For five iterations four parts will then be used for training and the remaining part for testing.
If test_size is a number between 0 and 1, in contrast, validation is done using a
shuffle-split
approach. Here, instead of splitting the data into n_splits equal parts up front, in each iteration
we randomize the data and sample a proportion equal to test_size to use for evaluation and the remaining
rows for training.
Whether to split the data by time.
Most recent data will be used for testing and previous data for training. Assumes data is passed
already sorted ascending by time.
One or more metrics/scoring functions to evaluate the model with.
When none is provided, will measure default metrics appropriate for the prediction task
(classification vs. regression determined from model or type of target column). See
sklearn model evaluation
for further details.
Which search strategy to use for optimization.
Grid search explores all possible combinations of parameters specified in params.
Randomized search, on the other hand, randomly samples iterations parameter combinations
from the distributions specified in params.
Configure model validation.
Allows evaluation of model performance via cross-validation
using custom metrics. If not specified, will by default perform 5-fold cross-validation with automatically selected
metrics.
Number of train-test splits to evaluate the model on.
Will split the dataset into training and test set n_splits times, train on the former
and evaluate on the latter using specified or automatically selected metrics.
What proportion of the data to use for testing in each split.
If null or not provided, will use k-fold cross-validation
to split the dataset. E.g. if n_splits is 5, the dataset will be split into 5 equal-sized parts.
For five iterations four parts will then be used for training and the remaining part for testing.
If test_size is a number between 0 and 1, in contrast, validation is done using a
shuffle-split
approach. Here, instead of splitting the data into n_splits equal parts up front, in each iteration
we randomize the data and sample a proportion equal to test_size to use for evaluation and the remaining
rows for training.
Whether to split the data by time.
Most recent data will be used for testing and previous data for training. Assumes data is passed
already sorted ascending by time.
One or more metrics/scoring functions to evaluate the model with.
When none is provided, will measure default metrics appropriate for the prediction task
(classification vs. regression determined from model or type of target column). See
sklearn model evaluation
for further details.
The parameter values to explore.
Allows tuning of any and all of the parameters that can be set also as constants in the
“params” attribute.
Keys in this object should be strings identifying parameter names, and values should be
lists of values to explore for that parameter. E.g. "depth": [3, 5, 7].
Number of iterations.
The maximum number of trees that can be built when solving machine learning problems. When using other parameters that limit the number of iterations, the final number of trees may be less than the number specified in this parameter.
Maximum cardinality of variables to be one-hot encoded.
Use one-hot encoding for all categorical features with a number of different values less than or equal to this value. Other variables will be target-encoded. Note that one-hot encoding is faster than the alternatives, so decreasing this value makes it more likely slower methods will be used. See CatBoost details for further information.
The maximum number of features that can be combined when transforming categorical variables.
Each resulting combination consists of one or more categorical features and can optionally contain binary features in the following form: “numeric feature > value”.
The amount of randomness to use for scoring splits.
Use this parameter to avoid overfitting the model. The value multiplies the variance of a random variable (with
zero mean) that is added to the score used to select splits when a tree is grown.
Maximum depth of the tree.
The maximum depth of the tree. If None, then nodes are expanded until all leaves are pure
or until all leaves contain less than min_samples_split samples.
Minimum number of samples required to split an internal node.
The minimum number of samples required to split an internal node. If int, then consider min_samples_split
as the minimum count. If float, then min_samples_split is a fraction and ceil(min_samples_split * n_samples)
are the minimum number of samples for each split.
Minimum number of samples required to be at a leaf node.
The minimum number of samples required to be at a leaf node. A split point at any depth will only be
considered if it leaves at least min_samples_leaf training samples in each of the left and right branches.
This may have the effect of smoothing the model, especially in regression.
If int, then consider min_samples_leaf as the minimum count. If float, then min_samples_leaf is
a fraction and ceil(min_samples_leaf * n_samples) are the minimum number of samples for each node.
Grow a tree with max_leaf_nodes in best-first fashion.
Best nodes are defined as relative reduction in impurity. If None then unlimited number of leaf nodes.
Number of features to consider when looking for the best split.
The number of features to consider when looking for the best split:
If int, then consider max_features features at each split.
If float, then max_features is a fraction and int(max_features * n_features) features are considered at each split.
If “auto”, then max_features=sqrt(n_features).
If “sqrt”, then max_features=sqrt(n_features).
If “log2”, then max_features=log2(n_features).
If None, then max_features=n_features.
Note: the search for a split does not stop until at least one valid partition of the node samples is found,
even if it requires to effectively inspect more than max_features features.
Complexity parameter used for Minimal Cost-Complexity Pruning.
Minimal Cost-Complexity Pruning recursively finds the node with the “weakest link”. The weakest link is
characterized by an effective alpha, where the nodes with the smallest effective alpha are pruned first.
As alpha increases, more of the tree is pruned, which increases the total impurity of its leaves.
Sort the data before training.
If the data is not already sorted by time, you can sort it here. This is useful when you want to split the data
by time, for example to train on older data and test on newer data (see the time_split parameter in validation
configurations). If the data is already sorted by time, you can ignore this parameter.
Sort order.
Whether to sort in ascending or descending order. If the single value true is provided, or no value is specified,
all columns will be sorted in ascending order. If a single false is provided, all columns will be sorted in descending
order. If an array of booleans is provided, each column will be sorted according to the corresponding boolean value.
