Skip to content

Data management

Data sets

ATOM is designed to work around one single dataset: the one with which atom is initialized. This is the dataset you want to explore, transform, and use for model training and validation. ATOM differentiates three different data sets:

  • The training set is usually the largest of the data sets. As the name suggests, this set is used to train the pipeline. During hyperparameter tuning, only the training set is used to fit and evaluate the estimator in every call. The training set in the current branch can be accessed through the train attribute. It's features and target can be accessed through X_train and y_train respectively.
  • The test set is used to evaluate the models. The model scores on this set give an indication on how the model performs on new data. The test set can be accessed through the test attribute. It's features and target can be accessed through X_test and y_test respectively.
  • The holdout set is an optional, separate set that should only be used to evaluate the final model's performance. Create this set when you are going to use the test set for an intermediate validation step. The holdout set is immediately set apart during initialization and is not considered part of atom's dataset (the dataset attribute only returns the train and test sets). The holdout set is left untouched until predictions are made on it, i.e., it does not undergo any pipeline transformations until the data set is requested for the first time. The holdout set is stored in atom's holdout attribute. See herean example that shows how to use the holdout data set.

The data can be provided in different formats. If the data sets are not specified beforehand, you can input the features and target separately or together:

  • X
  • X, y

Remember to use the y parameter to indicate the target column in X when using the first option. If not specified, the last column in X is used as the target. In both these cases, the sizes of the sets are defined using the test_size and holdout_size parameters. Note that the splits are made after the subsample of the dataset with the n_rows parameter (when not left to its default value).

If you already have the separate data sets, provide them using one of the following formats:

  • train, test
  • train, test, holdout
  • X_train, X_test, y_train, y_test
  • X_train, X_test, X_holdout, y_train, y_test, y_holdout
  • (X_train, y_train), (X_test, y_test)
  • (X_train, y_train), (X_test, y_test), (X_holdout, y_holdout)

The input data is always converted internally to a dataframe if it isn't one already. The column names should always be strings. If they are not, atom changes their type at initialization. If no column names are provided, default names are given of the form X[N-1], where N stands for the n-th feature in the dataset.


Indexing

By default, atom resets the dataframe's index after initialization and after every transformation in the pipeline. To avoid this, specify the index parameter. If the dataset has an 'identifier' column, it is useful to use it as index for two reasons:

  • An identifier doesn't usually contain any useful information on the target column, and should therefore be removed before training.
  • Predictions of specific rows can be accessed through their index.

Warning

Avoid duplicate indices in the dataframe. Having them raises an error when initializing atom and may potentially lead to unexpected behavior if introduced later.


Sparse datasets

If atom is initialized using a scipy sparse matrix, it is converted internally to a dataframe of sparse columns. Read more about pandas' sparse data structures here. The same conversion takes place when a transformer returns a sparse matrix, like for example, the Vectorizer.

Note that ATOM considers a dataset to be sparse if any of the columns is sparse. A dataset can only benefit from sparsity when all its columns are sparse, hence mixing sparse and non-sparse columns is not recommended and can cause estimators to decrease their training speed or even crash. Use the shrink method to convert dense features to sparse and the available_models method to check which models have native support for sparse matrices.

Click here to see an example that uses sparse data.

Warning

Estimators accelerated using sklearnex don't support sparse datasets.


Multioutput tasks

Multioutput is a task where there are more than one target column, i.e., the goal is to predict multiple targets at the same time. When providing a dataframe as target, use the y parameter. Providing y without keyword makes ATOM think you are providing train, test (see the data sets section).

Task types

ATOM recognizes four multioutput tasks.

Note

Combinations of binary and multiclass target columns are treated as multiclass-multioutput tasks.

Multilabel

Multilabel is a classification task, labeling each sample with m labels from n_classes possible classes, where m can be 0 to n_classes inclusive. This can be thought of as predicting properties of a sample that are not mutually exclusive.

For example, prediction of the topics relevant to a text document. The document may be about one of religion, politics, finance or education, several of the topic classes or all of the topic classes. The target column (atom.y) could look like this:

0                        [politics]
1               [religion, finance]
2    [politics, finance, education]
3                                []
4                         [finance]
5               [finance, religion]
6                         [finance]
7               [religion, finance]
8                       [education]
9     [finance, religion, politics]

Name: target, dtype: object

A model can not directly ingest a variable amount of target classes. Use the clean method to assign a binary output to each class, for every sample. Positive classes are indicated with 1 and negative classes with 0. It is thus comparable to running n_classes binary classification tasks. In our example, the target (atom.y) is converted to:

   education  finance  politics  religion
0          0        0         1         0
1          0        1         0         1
2          1        1         1         0
3          0        0         0         0
4          0        1         0         0
5          0        1         0         1
6          0        1         0         0
7          0        1         0         1
8          1        0         0         0
9          0        1         1         1

Multiclass-multioutput

Multiclass-multioutput (also known as multitask classification) is a classification task which labels each sample with a set of non-binary properties. Both the number of properties and the number of classes per property is greater than 2. A single estimator thus handles several joint classification tasks. This is both a generalization of the multilabel classification task, which only considers binary attributes, as well as a generalization of the multiclass classification task, where only one property is considered.

For example, classification of the properties "type of fruit" and "colour" for a set of images of fruit. The property "type of fruit" has the possible classes: "apple", "pear" and "orange". The property "colour" has the possible classes: "green", "red", "yellow" and "orange". Each sample is an image of a fruit, a label is output for both properties and each label is one of the possible classes of the corresponding property.

Multioutput regression

Multioutput regression predicts multiple numerical properties for each sample. Each property is a numerical variable and the number of properties to be predicted for each sample is >= 2. Some estimators that support multioutput regression are faster than just running n_output estimators.

For example, prediction of both wind speed and wind direction, in degrees, using data obtained at a certain location. Each sample would be data obtained at one location and both wind speed and direction would be output for each sample.

Multivariate

Multivariate is the multioutput task for forecasting. In this case, we try to forecast more than one time series at the same time.

Although all forecasting models in ATOM support multioutput tasks (thus have the native multioutput=True flag), we still differentiate two types of models:

  • The "genuine multioutput" models apply forecasts where every prediction of endogenous (y) variables will depend on values of the other target columns.
  • The rest of the models apply an estimator per column, meaning that forecasts will be made per endogenous variable, and not be affected by other variables. To access the column-wise estimators, use the estimator's forecasters_ parameter, which stores the fitted forecasters in a dataframe.

Tip

Use sktime's get_tags() method to check if an estimator is "genuine multioutput", e.g. atom.tbats.estimator.get_tags(). Search for the scitype:y key in the response. If the value is 'univariate', the estimator is genuine multioutput, and if 'multivariate', it isn't.

Read more about time series tasks here.

Native multioutput models

Some models have native support for multioutput tasks. This means that the original estimator is used to make predictions directly on all the target columns. Read in the model selection section how to get an overview of all models and their tags, including the native_multioutput.

Non-native multioutput models

The majority of the models don't have integrated support for multioutput tasks. However, it's possible to still use them for such tasks, wrapping them in a meta-estimator capable of handling multiple target columns. For non-native multioutput models, ATOM does so automatically. For multilabel tasks, the meta-estimator is:

And for multiclass-multioutput and multioutput regression, the meta-estimators are respectively:

Warning

Currently, scikit-learn metrics do not support multiclass-multioutput classification tasks. In this case, ATOM calculates the mean of the selected metric over every target.

Tip


Branches

You might want to compare how a model performs on a dataset transformed through multiple pipelines, each using different transformers. For example, on one pipeline with an undersampling strategy and the other with an oversampling strategy. To be able to do this, ATOM has a branching system.

The branching system helps the user to manage multiple data pipelines within the same atom instance. Branches are created and accessed through atom's branch property. A branch contains a specific pipeline, the dataset transformed through that pipeline, and all data and utility attributes that refer to that dataset. Transformers and models called from atom use the dataset in the current branch, as well as data attributes such as atom.dataset. It's not allowed to change the data in a branch after fitting a model with it. Instead, create a new branch for every unique pipeline.

By default, atom starts with one branch called "main". To start a new branch, set a new name to the property, e.g., atom.branch = "undersample". This creates a new branch from the current one. To create a branch from any other branch type "_from_" between the new name and the branch from which to split, e.g., atom.branch = "oversample_from_main" creates branch "oversample" from branch "main", even if the current branch is "undersample". To switch between existing branches, just type the name of the desired branch, e.g., atom.branch = "main" brings you back to the main branch. Note that every branch contains a unique copy of the whole dataset! Creating many branches can cause memory issues for large datasets.

See the Imbalanced datasets or Feature engineering examples for branching use cases.

Warning

Always create a new branch if you want to change the dataset after fitting a model! Forcing a data change through the data property's @setter can cause unexpected model behavior and break down the plotting methods.


diagram_branch

Figure 1. Diagram of a possible branch system to compare an oversampling with an undersampling pipeline.


Memory considerations

An atom instance stores one copy of the dataset for each branch (this doesn't include the holdout set, which is only stored once), and one copy of the initial dataset with which the instance is initialized. This copy of the original dataset is necessary to avoid data leakage during hyperparameter tuning and for some specific methods like cross_validate and reset. It's created as soon as there are no branches in the initial state (usually after calling the first data transformation). If the dataset is occupying too much memory, consider using the shrink method to convert the dtypes to their smallest possible matching dtype.

When working with large datasets and multiple branches, it becomes impossible to store all branches in memory at the same time. To avoid out-of-memory errors, use atom's memory parameter. If not False, atom saves the data of inactive branches as well as the original branch at the specified location (in a directory called joblib, the name of the underlying library managing the caching), maintaining only the current active branch in memory. This mechanism results in a slight drop in performance because of the I/O overhead, but can save a lot of memory. Additionally, the memory's location is also used to cache the output of the fit and transform methods of steps in the pipeline. See here an example using the memory parameter.

Apart from the dataset itself, a model's metric scores and shap values are also stored as attributes of the model to avoid having to recalculate them every time they are needed. You can delete all these attributes using the clear method in order to free some memory before saving atom.


Data transformations

Performing data transformations is a common requirement of many datasets before they are ready to be ingested by a model. ATOM provides various classes to apply data cleaning and feature engineering transformations to the data. This tooling should be able to help you apply most of the typically needed transformations to get the data ready for modeling. For further fine-tuning, it's also possible to transform the data using custom transformers (see the add method) or through a function (see the apply method). Remember that all transformations are only applied to the dataset in the current branch.


Row and column selection

Many methods in atom contain the rows or columns parameter to select a subset of the dataset. Examples are the evaluate and save_data methods for rows, and the distributions and shrink methods for columns. All data cleaning and feature engineering methods use the columns parameter to apply the transformation only to that selection of columns, and all prediction methods use the rows parameter to make predictions on that selection of rows.

As you can see, these two parameters are very important and shared across many methods in atom. Rows and columns can be selected in multiple ways. The check is performed in the order described hereunder:

  1. By actual dataset, e.g., rows=atom.test is equal to rows="test".
  2. By range or slice, e.g., rows=range(100) to select the first 100 rows from the dataset or rows=slice(20, 100) to select rows 20 to 99.
  3. By exact name, e.g., rows=["row1", "row2"] to select rows with indices row1 and row2 or columns=["col1", "col2"] to select columns col1 and col2. It's also possible to use the + sign to select multiple rows or columns, e.g., columns="col1+col2 is the same as columns=["col1", "col2"].
  4. By position, e.g., rows=[0, 1, 2] to select the first three rows.
  5. By name of the data set (only for rows), e.g., rows="train" to select all rows in the training set, or rows="test+holdout" to select all rows in the test and holdout sets. Valid data sets are dataset, train, test and holdout.
  6. By dtype (only for columns), e.g., columns="number" to select only numerical columns. See pandas' user guide.
  7. By regex match, e.g., columns="mean_.*" to select all columns starting with mean_.
  8. Excluding instead of including using the ! sign, e.g. columns="!col1" to select all columns except col1. You can also exclude multiple rows or columns like this columns=["!col1", "!col2"] or this columns="!col1+!col2". It's also possible to exclude data sets for row selection, e.g., columns="!train" or dtypes for column selection, e.g., columns="!number". Note that if a column name starts with !, the selection of that name will take priority over exclusion. Rows and columns can only be included or excluded, and not both at the same time. For example, this selection raises an exception column=["col1", "!col2"].

Additionally, the forecast horizon (parameter fh) in forecasting tasks can be selected much in the same way as rows, where the horizon is inferred as the index of the row selection. Note that, contrary to sktime's API but for consistency with the rest of ATOM's API, atom's fh starts with the training set, i.e., selecting atom.nf.predict(fh=range(5)) forecasts the first 5 rows of the training set, not the test set. To get the same result as sktime, use a ForecastingHorizon object, e.g., atom.nf.predict(fh=ForecastingHorizon(range(5))).

Info

In some plotting methods, it's possible to plot separate lines for different subsets of the rows. For example, to compare the results on the train and test set. For these cases, either provide a sequence to the rows parameter for every line you want to draw, e.g., atom.plot_roc(rows=("train", "test")), or provide a dictionary where the keys are the names of the sets (used in the legend) and the values are the corresponding selection of rows, selected using any of the aforementioned approaches, e.g, atom.plot_roc(rows={"0-99": range(100), "100-199": range(100, 200}). Note that for these methods, using atom.plot_roc(rows="train+test"), only plots one line with the data from both sets. See the advanced plotting example.


Data engines

ATOM is mostly built around sklearn (and sktime for time series tasks), and both these libraries use numpy as their computation backend. Since atom relies heavily on column names, it uses pandas (which in turn uses numpy) as its data backend. However, for the convenience of the user, it implements several data engines, that wraps the data in a different type when called by the user. This is very similar to sklearn's set_output behaviour, but ATOM extends this to many more data types. For example, selecting the polars data engine, makes atom.dataset return a polars dataframe and atom.winner.predict(X) return a polars series. See here an example notebook.

The data engine can be specified through the engine parameter, e.g. engine="pyarrow" or engine={"data": "pyarrow","estimator": "sklearnex"} to combine it with an [estimator engine][estimator acceleration]. ATOM integrates the following data engines:

  • numpy: Transform the data to a numpy array.
  • pandas: Leave the dataset as a pandas object. This is the default engine, that leaves the data unchanged.
  • pandas-pyarrow: Transform the data to pandas with the pyarrow backend. Read more in pandas' user guide.
  • polars: The polars library is a blazingly fast dataframe library implemented in Rust and based on Apache Arrow. Transforms the data to a polars dataframe or series.
  • polars-lazy: This engine is similar to the polars engine, but it returns a pl.LazyFrame instead of a pl.pd.DataFrame.
  • pyarrow: PyArrow is a cross-language, platform-independent, in-memory data format, that provides an efficient and fast way to serialize and deserialize data. the data is transformed to a pa.Table or pa.Array.
  • modin: The modin library is a multi-threading, drop-in replacement for pandas, that uses Ray as backend. Transform the data to a modin dataframe or series.
  • dask: The dask library is a powerful Python library for parallel and distributed computing. Transform the data to a dask dataframe or dask series.
  • pyspark: The pyspark library is the Python API for Apache Spark. Transform the data to a pyspark dataframe or pyspark series.
  • pyspark-pandas: Similar to the pyspark engine, but it returns pyspark objects with the pandas API.

Note

It's important to realize that, within atom, the data is still processed using pandas (with the numpy backend). Only when the data is returned to the user, it is transformed to the selected format.