evalml.model_understanding.partial_dependence¶
-
evalml.model_understanding.partial_dependence(pipeline, X, features, percentiles=(0.05, 0.95), grid_resolution=100, kind='average')[source]¶ Calculates one or two-way partial dependence. If a single integer or string is given for features, one-way partial dependence is calculated. If a tuple of two integers or strings is given, two-way partial dependence is calculated with the first feature in the y-axis and second feature in the x-axis.
- Parameters
pipeline (PipelineBase or subclass) – Fitted pipeline
X (pd.DataFrame, np.ndarray) – The input data used to generate a grid of values for feature where partial dependence will be calculated at
features (int, string, tuple[int or string]) – The target feature for which to create the partial dependence plot for. If features is an int, it must be the index of the feature to use. If features is a string, it must be a valid column name in X. If features is a tuple of int/strings, it must contain valid column integers/names in X.
percentiles (tuple[float]) – The lower and upper percentile used to create the extreme values for the grid. Must be in [0, 1]. Defaults to (0.05, 0.95).
grid_resolution (int) – Number of samples of feature(s) for partial dependence plot. If this value is less than the maximum number of categories present in categorical data within X, it will be set to the max number of categories + 1. Defaults to 100.
{'average' (kind) – The type of predictions to return. ‘individual’ will return the predictions for all of the points in the grid for each sample in X. ‘average’ will return the predictions for all of the points in the grid but averaged over all of the samples in X.
'individual' – The type of predictions to return. ‘individual’ will return the predictions for all of the points in the grid for each sample in X. ‘average’ will return the predictions for all of the points in the grid but averaged over all of the samples in X.
'both'} – The type of predictions to return. ‘individual’ will return the predictions for all of the points in the grid for each sample in X. ‘average’ will return the predictions for all of the points in the grid but averaged over all of the samples in X.
- Returns
When kind=’average’: DataFrame with averaged predictions for all points in the grid averaged over all samples of X and the values used to calculate those predictions.
When kind=’individual’: DataFrame with individual predictions for all points in the grid for each sample of X and the values used to calculate those predictions. If a two-way partial dependence is calculated, then the result is a list of DataFrames with each DataFrame representing one sample’s predictions.
When kind=’both’: A tuple consisting of the averaged predictions (in a DataFrame) over all samples of X and the individual predictions (in a list of DataFrames) for each sample of X.
In the one-way case: The dataframe will contain two columns, “feature_values” (grid points at which the partial dependence was calculated) and “partial_dependence” (the partial dependence at that feature value). For classification problems, there will be a third column called “class_label” (the class label for which the partial dependence was calculated). For binary classification, the partial dependence is only calculated for the “positive” class.
In the two-way case: The data frame will contain grid_resolution number of columns and rows where the index and column headers are the sampled values of the first and second features, respectively, used to make the partial dependence contour. The values of the data frame contain the partial dependence data for each feature value pair.
- Return type
pd.DataFrame, list(pd.DataFrame), or tuple(pd.DataFrame, list(pd.DataFrame))
- Raises
ValueError – if the user provides a tuple of not exactly two features.
ValueError – if the provided pipeline isn’t fitted.
ValueError – if the provided pipeline is a Baseline pipeline.
ValueError – if any of the features passed in are completely NaN
ValueError – if any of the features are low-variance. Defined as having one value occurring more than the upper percentile passed by the user. By default 95%.