ATOMClassifier
Main class for classification tasks.
Apply all data transformations and model management provided by the package on a given dataset. Note that, contrary to sklearn's API, the instance contains the dataset on which to perform the analysis. Calling a method will automatically apply it on the dataset it contains.
All data cleaning, feature engineering, model training and plotting functionality can be accessed from an instance of this class.
Parameters |
*arrays: sequence of indexables
Dataset containing features and target. Allowed formats are:
y: int, str, sequence or dataframe-like, default=-1
X, train, test: dataframe-like y: int, str, sequence or dataframe-like
Target column(s) corresponding to
index: bool, int, str or sequence, default=FalseX .
This parameter is ignored if the target column is provided
through
Handle the index in the resulting dataframe.
metadata: dict or None, default=None
Metadata to route to estimators, scorers, and CV splitters.
If None, no metadata is used. If dict, the available keys are:
ignore: int, str, sequence or None, default=None
Features in X to ignore during data transformations and model
training. The features are still used in the remaining methods.
test_size: int or float, default=0.2
This parameter is ignored if the test set is provided
through If 'groups' is provided in the
This parameter is ignored if the holdout set is provided
through
Whether to shuffle the dataset before splitting the data sets.
stratify: int, str or None, default=-1
Handle stratification of the target classes over the data sets.
n_rows: int or float, default=1
The stratification column can't contain This parameter is ignored if
Random subsample of the dataset to use. The default value selects
all rows.
n_jobs: int, default=1
Number of cores to use for parallel processing.
device: str, default="cpu"
Device on which to run the estimators. Use any string that
follows the SYCL_DEVICE_FILTER filter selector, e.g.
engine: str, dict or None, default=Nonedevice="gpu" to use the GPU. Read more in the
user guide.
Execution engine to use for data and
estimators. The value should be
one of the possible values to change one of the two engines,
or a dictionary with keys
backend: str, default="loky"data and estimator , with their
corresponding choice as values to change both engines. If
None, the default values are used. Choose from:
Parallelization backend. Read more in the
user guide. Choose from:
memory: bool, str, Path or Memory, default=False
Enables caching for memory optimization. Read more in the
user guide.
verbose: int, default=0
Verbosity level of the class. Choose from:
warnings: bool or str, default=False
Changing this parameter affects the
Name of the mlflow experiment to use for tracking.
If None, no mlflow tracking is performed.
random_state: int or None, default=None
Seed used by the random number generator. If None, the random
number generator is the RandomState used by np.random .
|
See Also
Example
>>> from atom import ATOMClassifier
>>> from sklearn.datasets import load_breast_cancer
>>> X, y = load_breast_cancer(return_X_y=True, as_frame=True)
>>> # Initialize atom
>>> atom = ATOMClassifier(X, y, verbose=2)
<< ================== ATOM ================== >>
Configuration ==================== >>
Algorithm task: Binary classification.
Dataset stats ==================== >>
Shape: (569, 31)
Train set size: 456
Test set size: 113
-------------------------------------
Memory: 138.97 kB
Scaled: False
Outlier values: 185 (1.3%)
>>> # Apply data cleaning and feature engineering methods
>>> atom.balance(strategy="smote")
Oversampling with SMOTE...
--> Adding 116 samples to class 0.
>>> atom.feature_selection(strategy="rfe", solver="lr", n_features=22)
Fitting FeatureSelector...
Performing feature selection ...
--> rfe selected 22 features from the dataset.
--> Dropping feature mean area (rank 4).
--> Dropping feature mean compactness (rank 3).
--> Dropping feature mean fractal dimension (rank 7).
--> Dropping feature smoothness error (rank 9).
--> Dropping feature concave points error (rank 5).
--> Dropping feature symmetry error (rank 2).
--> Dropping feature fractal dimension error (rank 8).
--> Dropping feature worst area (rank 6).
>>> # Train models
>>> atom.run(models=["LR", "RF", "XGB"])
Training ========================= >>
Models: LR, RF, XGB
Metric: f1
Results for LogisticRegression:
Fit ---------------------------------------------
Train evaluation --> f1: 0.9861
Test evaluation --> f1: 0.971
Time elapsed: 0.188s
-------------------------------------------------
Time: 0.188s
Results for RandomForest:
Fit ---------------------------------------------
Train evaluation --> f1: 1.0
Test evaluation --> f1: 0.971
Time elapsed: 0.180s
-------------------------------------------------
Time: 0.180s
Results for XGBoost:
Fit ---------------------------------------------
Train evaluation --> f1: 1.0
Test evaluation --> f1: 0.971
Time elapsed: 0.530s
-------------------------------------------------
Time: 0.530s
Final results ==================== >>
Total time: 0.908s
-------------------------------------
LogisticRegression --> f1: 0.971 !
RandomForest --> f1: 0.971 !
XGBoost --> f1: 0.971 !
>>> # Analyze the results
>>> atom.results
f1_train | f1_test | time_fit | time | |
---|---|---|---|---|
LR | 0.986100 | 0.971000 | 0.188171 | 0.188171 |
RF | 1.000000 | 0.971000 | 0.180164 | 0.180164 |
XGB | 1.000000 | 0.971000 | 0.530483 | 0.530483 |
Magic methods
The class contains some magic methods to help you access some of its elements faster. Note that methods that apply on the pipeline can return different results per branch.
- __repr__: Prints an overview of atom's branches, models, and metrics.
- __len__: Returns the length of the dataset.
- __iter__: Iterate over the pipeline's transformers.
- __contains__: Checks if the provided item is a column in the dataset.
- __getitem__: Access a branch, model, column or subset of the dataset.
Attributes
Data attributes
The data attributes are used to access the dataset and its properties. Updating the dataset will automatically update the response of these attributes accordingly.
Attributes |
pipeline: Pipeline Pipeline of transformers.
mapping: dict[str, dict[str, int | float]]Tip Use the plot_pipeline method to visualize the pipeline. Encoded values and their respective mapped values.
dataset: pd.DataFrameThe column name is the key to its mapping dictionary. Only for columns mapped to a single column (e.g., Ordinal, Leave-one-out, etc...). Complete data set.
train: pd.DataFrameTraining set.
test: pd.DataFrameTest set.
X: pd.DataFrameFeature set.
y: pd.Series | pd.DataFrameTarget column(s).
holdout: pd.DataFrame | NoneHoldout set.
X_train: pd.DataFrameThis data set is untransformed by the pipeline. Read more in the user guide. Features of the training set.
y_train: pd.Series | pd.DataFrameTarget column(s) of the training set.
X_test: pd.DataFrameFeatures of the test set.
y_test: pd.Series | pd.DataFrameTarget column(s) of the test set.
shape: tuple[int, int]Shape of the dataset (n_rows, n_columns).
columns: pd.IndexName of all the columns.
n_columns: intNumber of columns.
features: pd.IndexName of the features.
n_features: intNumber of features.
target: str | list[str]Name of the target column(s).
scaled: boolWhether the feature set is scaled.
duplicates: intA data set is considered scaled when it has mean~0 and std~1, or when there is a scaler in the pipeline. Categorical and binary columns (only zeros and ones) are excluded from the calculation. Sparse datasets always return False. Number of duplicate rows in the dataset.
nans: pd.SeriesColumns with the number of missing values in them.
n_nans: intThis property is unavailable for sparse datasets. Number of rows containing missing values.
numerical: pd.IndexThis property is unavailable for sparse datasets. Names of the numerical features in the dataset.
n_numerical: intNumber of numerical features in the dataset.
categorical: pd.IndexNames of the categorical features in the dataset.
n_categorical: intNumber of categorical features in the dataset.
outliers: pd.SeriesColumns in training set with number of outlier values.
n_outliers: intThis property is unavailable for sparse datasets. Number of samples in the training set containing outliers.
classes: pd.DataFrameThis property is unavailable for sparse datasets. Distribution of target classes per data set.
n_classes: int | pd.SeriesThis property is only available for classification tasks. Number of classes in the target column(s).
This property is only available for classification tasks. |
Utility attributes
The utility attributes are used to access information about the models in the instance after training.
Attributes |
pos_label: bool | int | float | str Positive label for binary/multilabel classification tasks.
metadata: BunchMetadata of the dataset.
ignore: tuple[str, ...]Read more in the user guide. Names of the ignored columns.
missing: list[Any]These columns aren't used in the transformer pipeline nor for model training. Values that are considered "missing".
branch: BranchThese values are used by the clean and impute methods. Default values are: None, NaN, NA, NaT, +inf, -inf, "", "?", "NA", "nan", "NaN", "NaT", "none", "None", "inf", "-inf". Note that None, NaN, NA, +inf and -inf are always considered missing since they are incompatible with sklearn estimators. Current active branch.
models: str | list[str] | NoneUse the property's Name of the model(s).
metric: str | list[str] | NoneName of the metric(s).
winners: list[model] | NoneModels ordered by performance.
winner: model | NonePerformance is measured as the highest score on the model's
Best performing model.
results: StylerPerformance is measured as the highest score on the model's
Overview of the training results.
All durations are in seconds. Possible values include:
Tip This attribute returns a pandas' Styler object. Convert
the result back to a regular dataframe using its |
Tracking attributes
The tracking attributes are used to customize what elements of the experiment are tracked. Read more in the user guide.
Plot attributes
The plot attributes are used to customize the plot's aesthetics. Read more in the user guide.
Attributes |
palette: str | Sequence[str] Color palette.
title_fontsize: int | floatSpecify one of plotly's built-in palettes or create
a custom one, e.g., Fontsize for the plot's title.
label_fontsize: int | floatFontsize for the labels, legend and hover information.
tick_fontsize: int | floatFontsize for the ticks along the plot's axes.
line_width: int | floatWidth of the line plots.
marker_size: int | floatSize of the markers.
|
Utility methods
Next to the plotting methods, the class contains a variety of utility methods to handle the data and manage the pipeline.
add | Add a transformer to the pipeline. |
apply | Apply a function to the dataset. |
available_models | Give an overview of the available predefined models. |
canvas | Create a figure with multiple plots. |
clear | Reset attributes and clear cache from all models. |
delete | Delete models. |
distributions | Get statistics on column distributions. |
eda | Create an Exploratory Data Analysis report. |
evaluate | Get all models' scores for the provided metrics. |
export_pipeline | Export the internal pipeline. |
get_class_weight | Return class weights for a balanced data set. |
get_sample_weight | Return sample weights for a balanced data set. |
inverse_transform | Inversely transform new data through the pipeline. |
load | Load an atom instance from a pickle file. |
merge | Merge another instance of the same class into this one. |
update_layout | Update the properties of the plot's layout. |
update_traces | Update the properties of the plot's traces. |
reset | Reset the instance to it's initial state. |
reset_aesthetics | Reset the plot aesthetics to their default values. |
save | Save the instance to a pickle file. |
save_data | Save the data in the current branch to a .csv file. |
shrink | Convert the columns to the smallest possible matching dtype. |
stacking | Add a Stacking model to the pipeline. |
stats | Display basic information about the dataset. |
status | Get an overview of the branches and models. |
transform | Transform new data through the pipeline. |
voting | Add a Voting model to the pipeline. |
Add a transformer to the pipeline.
If the transformer is not fitted, it is fitted on the complete training set. Afterward, the data set is transformed and the estimator is added to atom's pipeline. If the estimator is a sklearn Pipeline, every estimator is merged independently with atom.
Warning
- The transformer should have fit and/or transform methods
with arguments
X
(accepting a dataframe-like object of shape=(n_samples, n_features)) and/ory
(accepting a sequence of shape=(n_samples,)). - The transform method should return a feature set as a dataframe-like object of shape=(n_samples, n_features) and/or a target column as a sequence of shape=(n_samples,).
Note
If the transform method doesn't return a dataframe:
- The column naming happens as follows. If the transformer
has a
get_feature_names_out
method, it is used. If not, and it returns the same number of columns, the names are kept equal. If the number of columns changes, old columns will keep their name (as long as the column is unchanged) and new columns will receive the namex[N-1]
, where N stands for the n-th feature. This means that a transformer should only transform, add or drop columns, not combinations of these. - The index remains the same as before the transformation. This means that the transformer should not add, remove or shuffle rows unless it returns a dataframe.
Parameters |
transformer: Transformer
Estimator to add to the pipeline. Should implement a
columns: int, str, segment, sequence, dataframe or None, default=Nonetransform method. If a class is provided (instead of an
instance), and it has the n_jobs and/or random_state
parameters, it adopts atom's values.
Selection of columns to
transform. Only select features or the target column, not
both at the same time (if that happens, the target column
is ignored). If None, transform all columns.
train_only: bool, default=False
Whether to apply the estimator only on the training set or
on the complete dataset. Note that if True, the transformation
is skipped when making predictions on new data.
feature_names_out: "one-to-one", callable or None, default=None
Determines the list of feature names that will be returned
by the
**fit_paramsget_feature_names_out method.
Additional keyword arguments for the transformer's fit method.
|
Apply a function to the dataset.
This method is useful for stateless transformations such as taking the log, doing custom scaling, etc...
Note
This approach is preferred over changing the dataset directly
through the property's @setter
since the transformation is
stored in the pipeline.
Tip
Use atom.apply(lambda df: df.drop("column_name",axis=1))
to store the removal of columns in the pipeline.
Give an overview of the available predefined models.
Parameters |
**kwargs
Filter the returned models providing any of the column as
keyword arguments, where the value is the desired filter,
e.g., accepts_sparse=True , to get all models that accept
sparse input or supports_engines="cuml" to get all models
that support the cuML engine.
|
Returns |
pd.DataFrame
Tags of the available predefined models. The columns
depend on the task, but can include:
|
Create a figure with multiple plots.
This @contextmanager
allows you to draw many plots in one
figure. The default option is to add two plots side by side.
See the user guide for an example.
Parameters |
rows: int, default=1
Number of plots in length.
cols: int, default=2
Number of plots in width.
sharex: bool, default=False
If True, hide the label and ticks from non-border subplots
on the x-axis.
sharey: bool, default=False
If True, hide the label and ticks from non-border subplots
on the y-axis.
hspace: float, default=0.05
Space between subplot rows in normalized plot coordinates.
The spacing is relative to the figure's size.
vspace: float, default=0.07
Space between subplot cols in normalized plot coordinates.
The spacing is relative to the figure's size.
title: str, dict or None, default=None
Title for the plot.
legend: bool, str or dict, default="out"
Legend for the plot. See the user guide for
an extended description of the choices.
figsize: tuple or None, default=None
Figure's size in pixels, format as (x, y). If None, it
adapts the size to the number of plots in the canvas.
filename: str, Path or None, default=None
Save the plot using this name. Use "auto" for automatic
naming. The type of the file depends on the provided name
(.html, .png, .pdf, etc...). If
display: bool, default=Truefilename has no file type,
the plot is saved as html. If None, the plot is not saved.
Whether to render the plot.
|
Yields | {#canvas-go.Figure}
go.Figure
Plot object.
|
Reset attributes and clear cache from all models.
Reset certain model attributes to their initial state, deleting potentially large data arrays. Use this method to free some memory before saving the instance. The affected attributes are:
- In-training validation scores
- Shap values
- App instance
- Dashboard instance
- Calculated holdout data sets
Delete models.
If all models are removed, the metric is reset. Use this method to drop unwanted or to free some memory before saving. Deleted models are not removed from any active mlflow experiment.
Parameters |
models: int, str, Model, segment, sequence or None, default=None
Models to delete. If None, all models are deleted.
|
Get statistics on column distributions.
Compute the Kolmogorov-Smirnov test for various distributions against columns in the dataset. Only for numerical columns. Missing values are ignored.
Tip
Use the plot_distribution method to plot a column's distribution.
Parameters |
distributions: str, sequence or None, default=None
Names of the distributions in
columns: int, str, segment, sequence, dataframe or None, default=Nonescipy.stats to get the
statistics on. If None, a selection of the most common
ones is used.
Selection of columns on which
to perform the test. If None, select all numerical columns.
|
Returns |
pd.DataFrame
Statistic results with multiindex levels:
|
Create an Exploratory Data Analysis report.
ATOM uses the sweetviz package for EDA. The report is
rendered directly in the notebook. It can also be accessed
through the report
attribute. It can either report one
dataset or compare two datasets against each other.
Warning
This method can be slow for large datasets.
Parameters |
rows: str, sequence or dict, default="dataset"
Selection of rows on which to calculate the metric.
target: int or str, default=0
Target column to look at. Only for multilabel tasks. Only
bool and numerical features can be used as target.
filename: str, Path or None, default=None
Filename or pathlib.Path of the (html) file to save. If
None, don't save anything.
|
Get all models' scores for the provided metrics.
Tip
This method returns a pandas' Styler object. Convert
the result back to a regular dataframe using its data
attribute.
Parameters |
metric: str, func, scorer, sequence or None, default=None
Metric to calculate. If None, it returns an overview of
the most common metrics per task.
rows: hashable, segment, sequence or dataframe, default="test"
Selection of rows to calculate
metric on.
|
Returns | {#evaluate-Styler}
Styler
Scores of the models.
|
Export the internal pipeline.
This method returns a deepcopy of the branch's pipeline. Optionally, you can add a model as final estimator. The returned pipeline is already fitted on the training set.
Parameters |
model: str, Model or None, default=None
Model for which to export the pipeline. If the model used
automated feature scaling, the Scaler is added to
the pipeline. If None, the pipeline in the current branch
is exported (without any model).
|
Returns | {#export_pipeline-Pipeline}
Pipeline
Current branch as a sklearn-like Pipeline object.
|
Return class weights for a balanced data set.
Statistically, the class weights re-balance the data set so that the sampled data set represents the target population as closely as possible. The returned weights are inversely proportional to the class frequencies in the selected rows.
Parameters |
rows: hashable, segment, sequence or dataframe, default="train"
Selection of rows for which to
get the weights.
|
Returns |
dict
Classes with the corresponding weights. A dict of dicts is
returned for multioutput tasks.
|
Return sample weights for a balanced data set.
The returned weights are inversely proportional to the class
frequencies in the selected data set. For multioutput tasks,
the weights of each column of y
will be multiplied.
Parameters |
rows: hashable, segment, sequence or dataframe, default="train"
Selection of rows for which to
get the weights.
|
Returns |
pd.Series
Sequence of weights with shape=(n_samples,).
|
Inversely transform new data through the pipeline.
Transformers that are only applied on the training set are
skipped. The rest should all implement an inverse_transform
method. If only X
or only y
is provided, it ignores
transformers that require the other parameter. This can be
used to transform only the target column.
Load an atom instance from a pickle file.
If the instance was saved using save_data=False
,
it's possible to load new data into it and apply all data
transformations.
Info
The loaded instance's current branch is the same branch as it was when saved.
Parameters |
filename: str or Path
Filename or pathlib.Path of the pickle file.
data: tuple of indexables or None, default=None
Original dataset as it was provided to the instance's
constructor. Only use this parameter if the loaded file
was saved using save_data=False . Allowed formats are:
X, train, test: dataframe-like y: int, str, sequence or dataframe
|
Returns |
atom
Unpickled atom instance.
|
Merge another instance of the same class into this one.
Branches, models, metrics and attributes of the other instance
are merged into this one. If there are branches and/or models
with the same name, they are merged adding the suffix
parameter to their name. The errors and missing attributes are
extended with those of the other instance. It's only possible
to merge two instances if they are initialized with the same
dataset and trained with the same metric.
Update the properties of the plot's layout.
Recursively update the structure of the original layout with the values in the arguments.
Parameters |
**kwargs
Keyword arguments for the figure's update_layout method.
|
Update the properties of the plot's traces.
Recursively update the structure of the original traces with the values in the arguments.
Parameters |
**kwargs
Keyword arguments for the figure's update_traces method.
|
Reset the instance to it's initial state.
Deletes all branches and models. The dataset is also reset to its form after initialization.
Parameters |
hard: bool, default=False
If True, flushes completely the cache.
|
Reset the plot aesthetics to their default values.
Save the instance to a pickle file.
Parameters |
filename: str or Path, default="auto"
Filename or pathlib.Path of the file to save. Use
"auto" for automatic naming.
save_data: bool, default=True
Whether to save the dataset with the instance. This
parameter is ignored if the method is not called from atom.
If False, add the data to the load
method to reload the instance.
|
Save the data in the current branch to a .csv
file.
Parameters |
filename: str or Path, default="auto"
Filename or pathlib.Path of the file to save. Use
"auto" for automatic naming.
rows: hashable, segment, sequence or dataframe, default="dataset"
Selection of rows to save.
**kwargs
Additional keyword arguments for pandas' to_csv method.
|
Convert the columns to the smallest possible matching dtype.
Examples are: float64 -> float32, int64 -> int8, etc... Sparse arrays also transform their non-fill value. Use this method for memory optimization before saving the dataset. Note that applying transformers to the data may alter the types again.
Parameters |
int2bool: bool, default=False
Whether to convert
int2uint: bool, default=Falseint columns to bool type. Only if the
values in the column are strictly in (0, 1) or (-1, 1).
Whether to convert
str2cat: bool, default=Falseint to uint (unsigned integer). Only if
the values in the column are strictly positive.
Whether to convert
dense2sparse: bool, default=Falsestring to category . Only if the
number of categories is less than 30% of the column's length.
Whether to convert all features to sparse format. The value
that is compressed is the most frequent value in the column.
columns: int, str, segment, sequence, dataframe or None, default=None
Selection of columns to shrink. If
None, transform all columns.
|
Add a Stacking model to the pipeline.
Warning
Combining models trained on different branches into one ensemble is not allowed and will raise an exception.
Parameters |
models: segment, sequence or None, default=None
Models that feed the stacking estimator. The models must
have been fitted on the current branch.
name: str, default="Stack"
Name of the model. The name is always presided with the
model's acronym:
train_on_test: bool, default=FalseStack .
Whether to train the final estimator of the stacking model
on the test set instead of the training set. Note that
training it on the training set (default option) means there
is a high risk of overfitting. It's recommended to use this
option if you have another, independent set for testing
(holdout set).
**kwargs
Additional keyword arguments for one of these estimators.
Tip The model's acronyms can be used for the |
Display basic information about the dataset.
Get an overview of the branches and models.
This method prints the same information as the __repr__ and also saves it to the logger.
Transform new data through the pipeline.
Transformers that are only applied on the training set are
skipped. If only X
or only y
is provided, it ignores
transformers that require the other parameter. This can be
of use to, for example, transform only the target column.
Add a Voting model to the pipeline.
Warning
Combining models trained on different branches into one ensemble is not allowed and will raise an exception.
Parameters |
models: segment, sequence or None, default=None
Models that feed the stacking estimator. The models must have
been fitted on the current branch.
name: str, default="Vote"
Name of the model. The name is always presided with the
model's acronym:
**kwargsVote .
Additional keyword arguments for one of these estimators.
|
Data cleaning
The data cleaning methods can help you scale the data, handle missing values, categorical columns, outliers and unbalanced datasets. All attributes of the data cleaning classes are attached to atom after running. Read more in the user guide.
Tip
Use the eda method to examine the data and help you determine suitable parameters for the data cleaning methods.
balance | Balance the number of rows per class in the target column. |
clean | Apply standard data cleaning steps on the dataset. |
discretize | Bin continuous data into intervals. |
encode | Perform encoding of categorical features. |
impute | Handle missing values in the dataset. |
normalize | Transform the data to follow a Normal/Gaussian distribution. |
prune | Prune outliers from the training set. |
scale | Scale the data. |
Balance the number of rows per class in the target column.
When oversampling, the newly created samples have an increasing integer index for numerical indices, and an index of the form [estimator]_N for non-numerical indices, where N stands for the N-th sample in the data set.
See the Balancer class for a description of the parameters.
Warning
- The balance method does not support multioutput tasks.
- The balance method does not support
sample_weights
passed through metadata routing. - This transformation is only applied to the training set to maintain the original distribution of target classes in the test set.
Tip
Use atom's classes attribute for an overview of the target class distribution per data set.
Apply standard data cleaning steps on the dataset.
Use the parameters to choose which transformations to perform. The available steps are:
- Convert dtypes to the best possible types.
- Drop columns with specific data types.
- Remove characters from column names.
- Strip categorical features from spaces.
- Drop duplicate rows.
- Drop rows with missing values in the target column.
- Encode the target column (only for classification tasks).
See the Cleaner class for a description of the parameters.
Bin continuous data into intervals.
For each feature, the bin edges are computed during fit and, together with the number of bins, they will define the intervals. Ignores numerical columns.
See the Discretizer class for a description of the parameters.
Tip
Use the plot_distribution method to visualize a column's distribution and decide on the bins.
Perform encoding of categorical features.
The encoding type depends on the number of classes in the column:
- If n_classes=2 or ordinal feature, use Ordinal-encoding.
- If 2 < n_classes <=
max_onehot
, use OneHot-encoding. - If n_classes >
max_onehot
, usestrategy
-encoding.
Missing values are propagated to the output column. Unknown classes encountered during transforming are imputed according to the selected strategy. Rare classes can be replaced with a value in order to prevent too high cardinality.
See the Encoder class for a description of the parameters.
Note
This method only encodes the categorical features. It does not encode the target column! Use the clean method for that.
Tip
Use the categorical attribute for a list of the categorical features in the dataset.
Handle missing values in the dataset.
Impute or remove missing values according to the selected strategy. Also removes rows and columns with too many missing values.
See the Imputer class for a description of the parameters.
Tip
Transform the data to follow a Normal/Gaussian distribution.
This transformation is useful for modeling issues related to heteroscedasticity (non-constant variance), or other situations where normality is desired. Missing values are disregarded in fit and maintained in transform. Ignores categorical columns.
See the Normalizer class for a description of the parameters.
Tip
Use the plot_distribution method to examine a column's distribution.
Prune outliers from the training set.
Replace or remove outliers. The definition of outlier depends on the selected strategy and can greatly differ from one another. Ignores categorical columns.
See the Pruner class for a description of the parameters.
Note
This transformation is only applied to the training set in order to maintain the original distribution of samples in the test set.
Tip
Use the outliers attribute to check the number of outliers per column.
Scale the data.
Apply one of sklearn's scaling strategies. Categorical columns are ignored.
See the Scaler class for a description of the parameters.
Tip
Use the scaled attribute to check whether the dataset is scaled.
NLP
The Natural Language Processing (NLP) transformers help to convert raw
text to meaningful numeric values, ready to be ingested by a model. All
transformations are applied only on the column in the dataset called
corpus
. Read more in the user guide.
textclean | Apply standard text cleaning to the corpus. |
textnormalize | Normalize the corpus. |
tokenize | Tokenize the corpus. |
vectorize | Vectorize the corpus. |
Apply standard text cleaning to the corpus.
Transformations include normalizing characters and drop
noise from the text (emails, HTML tags, URLs, etc...). The
transformations are applied on the column named corpus
, in
the same order the parameters are presented. If there is no
column with that name, an exception is raised.
See the TextCleaner class for a description of the parameters.
Normalize the corpus.
Convert words to a more uniform standard. The transformations
are applied on the column named corpus
, in the same order the
parameters are presented. If there is no column with that name,
an exception is raised. If the provided documents are strings,
words are separated by spaces.
See the TextNormalizer class for a description of the parameters.
Tokenize the corpus.
Convert documents into sequences of words. Additionally,
create n-grams (represented by words united with underscores,
e.g., "New_York") based on their frequency in the corpus. The
transformations are applied on the column named corpus
. If
there is no column with that name, an exception is raised.
See the Tokenizer class for a description of the parameters.
Vectorize the corpus.
Transform the corpus into meaningful vectors of numbers. The
transformation is applied on the column named corpus
. If
there is no column with that name, an exception is raised.
If strategy="bow" or "tfidf", the transformed columns are named
after the word they are embedding with the prefix corpus_
. If
strategy="hashing", the columns are named hash[N], where N stands
for the n-th hashed column.
See the Vectorizer class for a description of the parameters.
Feature engineering
To further pre-process the data, it's possible to extract features from datetime columns, create new non-linear features transforming the existing ones, group similar features or, if the dataset is too large, remove features. Read more in the user guide.
feature_extraction | Extract features from datetime columns. |
feature_generation | Generate new features. |
feature_grouping | Extract statistics from similar features. |
feature_selection | Reduce the number of features in the data. |
Extract features from datetime columns.
Create new features extracting datetime elements (day, month,
year, etc...) from the provided columns. Columns of dtype
datetime64
are used as is. Categorical columns that can be
successfully converted to a datetime format (less than 30% NaT
values after conversion) are also used.
See the FeatureExtractor class for a description of the parameters.
Generate new features.
Create new combinations of existing features to capture the non-linear relations between the original features.
See the FeatureGenerator class for a description of the parameters.
Extract statistics from similar features.
Replace groups of features with related characteristics with new
features that summarize statistical properties of the group. The
statistical operators are calculated over every row of the group.
The group names and features can be accessed through the groups
method.
See the FeatureGrouper class for a description of the parameters.
Tip
Use a regex pattern with the groups
parameter to select
groups easier, e.g., atom.feature_grouping({"group1": "var_.+")
to select all features that start with var_
.
Reduce the number of features in the data.
Apply feature selection or dimensionality reduction, either to improve the estimators' accuracy or to boost their performance on very high-dimensional datasets. Additionally, remove multicollinear and low-variance features.
See the FeatureSelector class for a description of the parameters.
Note
- When strategy="univariate" and solver=None, f_classif or f_regression is used as default solver.
- When strategy is "sfs", "rfecv" or any of the advanced strategies and no scoring is specified, atom's metric (if it exists) is used as scoring.
Training
The training methods are where the models are fitted to the data and their performance is evaluated against a selected metric. There are three methods to call the three different training approaches. Read more in the user guide.
run | Train and evaluate the models in a direct fashion. |
successive_halving | Fit the models in a successive halving fashion. |
train_sizing | Train and evaluate the models in a train sizing fashion. |
Train and evaluate the models in a direct fashion.
Contrary to successive_halving and train_sizing, the direct approach only iterates once over the models, using the full dataset.
The following steps are applied to every model:
- Apply hyperparameter tuning (optional).
- Fit the model on the training set using the best combination of hyperparameters found.
- Evaluate the model on the test set.
- Train the estimator on various bootstrapped samples of the training set and evaluate again on the test set (optional).
See the DirectClassifier or DirectRegressor class for a description of the parameters.
Fit the models in a successive halving fashion.
The successive halving technique is a bandit-based algorithm that fits N models to 1/N of the data. The best half are selected to go to the next iteration where the process is repeated. This continues until only one model remains, which is fitted on the complete dataset. Beware that a model's performance can depend greatly on the amount of data on which it is trained. For this reason, it is recommended to only use this technique with similar models, e.g., only using tree-based models.
The following steps are applied to every model (per iteration):
- Apply hyperparameter tuning (optional).
- Fit the model on the training set using the best combination of hyperparameters found.
- Evaluate the model on the test set.
- Train the estimator on various bootstrapped samples of the training set and evaluate again on the test set (optional).
See the SuccessiveHalvingClassifier or SuccessiveHalvingRegressor class for a description of the parameters.
Train and evaluate the models in a train sizing fashion.
When training models, there is usually a trade-off between model performance and computation time; that is regulated by the number of samples in the training set. This method can be used to create insights in this trade-off, and help determine the optimal size of the training set. The models are fitted multiple times, ever-increasing the number of samples in the training set.
The following steps are applied to every model (per iteration):
- Apply hyperparameter tuning (optional).
- Fit the model on the training set using the best combination of hyperparameters found.
- Evaluate the model on the test set.
- Train the estimator on various bootstrapped samples of the training set and evaluate again on the test set (optional).
See the TrainSizingClassifier or TrainSizingRegressor class for a description of the parameters.