"""
A simple library of functions that provide feature importances
for scikit-learn random forest regressors and classifiers.
MIT License
Terence Parr, http://parrt.cs.usfca.edu
Kerem Turgutlu, https://www.linkedin.com/in/kerem-turgutlu-12906b65
"""
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from sklearn.ensemble import RandomForestClassifier
from sklearn.ensemble import RandomForestRegressor
from sklearn.ensemble.forest import _generate_unsampled_indices
from sklearn.ensemble import forest
from sklearn.model_selection import cross_val_score
from sklearn.base import clone
from sklearn.metrics import r2_score
from scipy.stats import spearmanr
from pandas.api.types import is_numeric_dtype
from matplotlib.colors import ListedColormap
from copy import copy
import warnings
def importances(model, X_valid, y_valid, features=None, n_samples=5000,
sort=True, metric=None, sample_weights=None):
"""
Compute permutation feature importances for scikit-learn models using
a validation set.
Given a Classifier or Regressor in model
and validation X and y data, return a data frame with columns
Feature and Importance sorted in reverse order by importance.
The validation data is needed to compute model performance
measures (accuracy or R^2). The model is not retrained.
You can pass in a list with a subset of features interesting to you.
All unmentioned features will be grouped together into a single
meta-feature on the graph. You can also pass in a list that has sublists
like:
[['latitude', 'longitude'], 'price', 'bedrooms']. Each string or sublist
will be permuted together as a feature or meta-feature; the drop in
overall accuracy of the model is the relative importance.
The model.score() method is called to measure accuracy drops.
This version that computes accuracy drops with the validation set
is much faster than the OOB, cross validation, or drop column
versions. The OOB version is a less vectorized because it needs to dig
into the trees to get out of examples. The cross validation and drop column
versions need to do retraining and are necessarily much slower.
This function used OOB not validation sets in 1.0.5; switched to faster
test set version for 1.0.6. (breaking API change)
:param model: The scikit model fit to training data
:param X_valid: Data frame with feature vectors of the validation set
:param y_valid: Series with target variable of validation set
:param features: The list of features to show in importance graph.
These can be strings (column names) or lists of column
names. E.g., features = ['bathrooms', ['latitude', 'longitude']].
Feature groups can overlap, with features appearing in multiple.
:param n_samples: How many records of the validation set to use
to compute permutation importance. The default is
5000, which we arrived at by experiment over a few data sets.
As we cannot be sure how all data sets will react,
you can pass in whatever sample size you want. Pass in -1
to mean entire validation set. Our experiments show that
not too many records are needed to get an accurate picture of
feature importance.
:param sort: Whether to sort the resulting importances
:param metric: Metric in the form of callable(model, X_valid, y_valid, sample_weights) to evaluate for,
if not set default's to model.score()
:param sample_weights: set if a different weighting is required for the validation samples
return: A data frame with Feature, Importance columns
SAMPLE CODE
rf = RandomForestRegressor(n_estimators=100, n_jobs=-1)
X_train, y_train = ..., ...
X_valid, y_valid = ..., ...
rf.fit(X_train, y_train)
imp = importances(rf, X_valid, y_valid)
"""
def flatten(features):
all_features = set()
for sublist in features:
if isinstance(sublist, str):
all_features.add(sublist)
else:
for item in sublist:
all_features.add(item)
return all_features
if not features:
# each feature in its own group
features = X_valid.columns.values
else:
req_feature_set = flatten(features)
model_feature_set = set(X_valid.columns.values)
# any features left over?
other_feature_set = model_feature_set.difference(req_feature_set)
if len(other_feature_set) > 0:
# if leftovers, we need group together as single new feature
features.append(list(other_feature_set))
X_valid, y_valid = sample(X_valid, y_valid, n_samples)
X_valid = X_valid.copy(deep=False) # we're modifying columns
baseline = None
if callable(metric):
baseline = metric(model, X_valid, y_valid, sample_weights)
else:
baseline = model.score(X_valid, y_valid, sample_weights)
imp = []
m = None
for group in features:
if isinstance(group, str):
save = X_valid[group].copy()
X_valid[group] = np.random.permutation(X_valid[group])
if callable(metric):
m = metric(model, X_valid, y_valid, sample_weights)
else:
m = model.score(X_valid, y_valid, sample_weights)
X_valid[group] = save
else:
save = {}
for col in group:
save[col] = X_valid[col].copy()
for col in group:
X_valid[col] = np.random.permutation(X_valid[col])
if callable(metric):
m = metric(model, X_valid, y_valid, sample_weights)
else:
m = model.score(X_valid, y_valid, sample_weights)
for col in group:
X_valid[col] = save[col]
imp.append(baseline - m)
# Convert and groups/lists into string column names
labels = []
for col in features:
if isinstance(col, list):
labels.append('\n'.join(col))
else:
labels.append(col)
I = pd.DataFrame(data={'Feature': labels, 'Importance': np.array(imp)})
I = I.set_index('Feature')
if sort:
I = I.sort_values('Importance', ascending=False)
return I
def sample(X_valid, y_valid, n_samples):
if n_samples < 0: n_samples = len(X_valid)
n_samples = min(n_samples, len(X_valid))
if n_samples < len(X_valid):
ix = np.random.choice(len(X_valid), n_samples)
X_valid = X_valid.iloc[ix].copy(deep=False) # shallow copy
y_valid = y_valid.iloc[ix].copy(deep=False)
return X_valid, y_valid
def sample_rows(X, n_samples):
if n_samples < 0: n_samples = len(X)
n_samples = min(n_samples, len(X))
if n_samples < len(X):
ix = np.random.choice(len(X), n_samples)
X = X.iloc[ix].copy(deep=False) # shallow copy
return X
def oob_importances(rf, X_train, y_train, n_samples=5000):
"""
Compute permutation feature importances for scikit-learn
RandomForestClassifier or RandomForestRegressor in arg rf.
Given training X and y data, return a data frame with columns
Feature and Importance sorted in reverse order by importance.
The training data is needed to compute out of bag (OOB)
model performance measures (accuracy or R^2). The model
is not retrained.
By default, sample up to 5000 observations to compute feature importances.
return: A data frame with Feature, Importance columns
SAMPLE CODE
rf = RandomForestRegressor(n_estimators=100, n_jobs=-1, oob_score=True)
X_train, y_train = ..., ...
rf.fit(X_train, y_train)
imp = oob_importances(rf, X_train, y_train)
"""
if isinstance(rf, RandomForestClassifier):
return permutation_importances(rf, X_train, y_train, oob_classifier_accuracy, n_samples)
elif isinstance(rf, RandomForestRegressor):
return permutation_importances(rf, X_train, y_train, oob_regression_r2_score, n_samples)
return None
def cv_importances(model, X_train, y_train, k=3):
"""
Compute permutation feature importances for scikit-learn models using
k-fold cross-validation (default k=3).
Given a Classifier or Regressor in model
and training X and y data, return a data frame with columns
Feature and Importance sorted in reverse order by importance.
Cross-validation observations are taken from X_train, y_train.
The model.score() method is called to measure accuracy drops.
return: A data frame with Feature, Importance columns
SAMPLE CODE
rf = RandomForestRegressor(n_estimators=100, n_jobs=-1)
X_train, y_train = ..., ...
rf.fit(X_train, y_train)
imp = cv_importances(rf, X_train, y_train)
"""
def score(model):
cvscore = cross_val_score(
model, # which model to use
X_train, y_train, # what training data to split up
cv=k) # number of folds/chunks
return np.mean(cvscore)
X_train = X_train.copy(deep=False) # shallow copy
baseline = score(model)
imp = []
for col in X_train.columns:
save = X_train[col].copy()
X_train[col] = np.random.permutation(X_train[col])
m = score(model)
X_train[col] = save
imp.append(baseline - m)
I = pd.DataFrame(data={'Feature': X_train.columns, 'Importance': np.array(imp)})
I = I.set_index('Feature')
I = I.sort_values('Importance', ascending=False)
return I
def permutation_importances(rf, X_train, y_train, metric, n_samples=5000):
imp = permutation_importances_raw(rf, X_train, y_train, metric, n_samples)
I = pd.DataFrame(data={'Feature':X_train.columns, 'Importance':imp})
I = I.set_index('Feature')
I = I.sort_values('Importance', ascending=False)
return I
def dropcol_importances(rf, X_train, y_train, metric=None, X_valid = None, y_valid = None, sample_weights = None):
"""
Compute drop-column feature importances for scikit-learn.
Given a RandomForestClassifier or RandomForestRegressor in rf
and training X and y data, return a data frame with columns
Feature and Importance sorted in reverse order by importance.
A clone of rf is trained once to get the baseline score and then
again, once per feature to compute the drop in either the model's .score() output
or a custom metric callable in the form of metric(model, X_valid, y_valid). In case of a custom metric
the X_valid and y_valid parameters should be set.
return: A data frame with Feature, Importance columns
SAMPLE CODE
rf = RandomForestRegressor(n_estimators=100, n_jobs=-1, oob_score=True)
X_train, y_train = ..., ...
rf.fit(X_train, y_train)
imp = dropcol_importances(rf, X_train, y_train)
"""
rf_ = clone(rf)
rf_.random_state = 999
rf_.fit(X_train, y_train)
baseline = rf_.oob_score_
imp = []
for col in X_train.columns:
X = X_train.drop(col, axis=1)
rf_ = clone(rf)
rf_.random_state = 999
rf_.fit(X, y_train)
if callable(metric):
o = metric(rf_, X_valid, y_valid, sample_weights)
else:
o = rf_.score(X_valid, y_valid, sample_weights)
imp.append(baseline - o)
imp = np.array(imp)
I = pd.DataFrame(data={'Feature':X_train.columns, 'Importance':imp})
I = I.set_index('Feature')
I = I.sort_values('Importance', ascending=False)
return I
def oob_dropcol_importances(rf, X_train, y_train):
"""
Compute drop-column feature importances for scikit-learn.
Given a RandomForestClassifier or RandomForestRegressor in rf
and training X and y data, return a data frame with columns
Feature and Importance sorted in reverse order by importance.
A clone of rf is trained once to get the baseline score and then
again, once per feature to compute the drop in out of bag (OOB)
score.
return: A data frame with Feature, Importance columns
SAMPLE CODE
rf = RandomForestRegressor(n_estimators=100, n_jobs=-1, oob_score=True)
X_train, y_train = ..., ...
rf.fit(X_train, y_train)
imp = oob_dropcol_importances(rf, X_train, y_train)
"""
rf_ = clone(rf)
rf_.random_state = 999
rf_.fit(X_train, y_train)
baseline = rf_.oob_score_
imp = []
for col in X_train.columns:
X = X_train.drop(col, axis=1)
rf_ = clone(rf)
rf_.random_state = 999
rf_.fit(X, y_train)
o = rf_.oob_score_
imp.append(baseline - o)
imp = np.array(imp)
I = pd.DataFrame(data={'Feature':X_train.columns, 'Importance':imp})
I = I.set_index('Feature')
I = I.sort_values('Importance', ascending=False)
return I
def importances_raw(rf, X_train, y_train, n_samples=5000):
if isinstance(rf, RandomForestClassifier):
return permutation_importances_raw(rf, X_train, y_train, oob_classifier_accuracy, n_samples)
elif isinstance(rf, RandomForestRegressor):
return permutation_importances_raw(rf, X_train, y_train, oob_regression_r2_score, n_samples)
return None
def permutation_importances_raw(rf, X_train, y_train, metric, n_samples=5000):
"""
Return array of importances from pre-fit rf; metric is function
that measures accuracy or R^2 or similar. This function
works for regressors and classifiers.
"""
X_sample, y_sample = sample(X_train, y_train, n_samples)
if not hasattr(rf, 'estimators_'):
rf.fit(X_sample, y_sample)
baseline = metric(rf, X_sample, y_sample)
X_train = X_sample.copy(deep=False) # shallow copy
y_train = y_sample
imp = []
for col in X_train.columns:
save = X_train[col].copy()
X_train[col] = np.random.permutation(X_train[col])
m = metric(rf, X_train, y_train)
X_train[col] = save
imp.append(baseline - m)
return np.array(imp)
def oob_classifier_accuracy(rf, X_train, y_train):
"""
Compute out-of-bag (OOB) accuracy for a scikit-learn random forest
classifier. We learned the guts of scikit's RF from the BSD licensed
code:
https://github.com/scikit-learn/scikit-learn/blob/a24c8b46/sklearn/ensemble/forest.py#L425
"""
X = X_train.values
y = y_train.values
n_samples = len(X)
n_classes = len(np.unique(y))
predictions = np.zeros((n_samples, n_classes))
for tree in rf.estimators_:
unsampled_indices = _generate_unsampled_indices(tree.random_state, n_samples)
tree_preds = tree.predict_proba(X[unsampled_indices, :])
predictions[unsampled_indices] += tree_preds
predicted_class_indexes = np.argmax(predictions, axis=1)
predicted_classes = [rf.classes_[i] for i in predicted_class_indexes]
oob_score = np.mean(y == predicted_classes)
return oob_score
def oob_regression_r2_score(rf, X_train, y_train):
"""
Compute out-of-bag (OOB) R^2 for a scikit-learn random forest
regressor. We learned the guts of scikit's RF from the BSD licensed
code:
https://github.com/scikit-learn/scikit-learn/blob/a24c8b46/sklearn/ensemble/forest.py#L702
"""
X = X_train.values if isinstance(X_train, pd.DataFrame) else X_train
y = y_train.values if isinstance(y_train, pd.Series) else y_train
n_samples = len(X)
predictions = np.zeros(n_samples)
n_predictions = np.zeros(n_samples)
for tree in rf.estimators_:
unsampled_indices = _generate_unsampled_indices(tree.random_state, n_samples)
tree_preds = tree.predict(X[unsampled_indices, :])
predictions[unsampled_indices] += tree_preds
n_predictions[unsampled_indices] += 1
if (n_predictions == 0).any():
warnings.warn("Too few trees; some variables do not have OOB scores.")
n_predictions[n_predictions == 0] = 1
predictions /= n_predictions
oob_score = r2_score(y, predictions)
return oob_score
def plot_importances(df_importances, save=None, xrot=0, tickstep=3,
label_fontsize=12,
figsize=None, scalefig=(1.0, 1.0), show=True):
"""
Given an array or data frame of importances, plot a horizontal bar chart
showing the importance values.
:param df_importances: A data frame with Feature, Importance columns
:type df_importances: pd.DataFrame
:param save: A filename identifying where to save the image.
:param xrot: Degrees to rotate importance (X axis) labels
:type xrot: int
:param tickstep: How many ticks to skip in X axis
:type tickstep: int
:param label_fontsize: The font size for the column names and x ticks
:type label_fontsize: int
:param figsize: Specify width and height of image (width,height)
:type figsize: 2-tuple of floats
:param scalefig: Scale width and height of image (widthscale,heightscale)
:type scalefig: 2-tuple of floats
:param show: Execute plt.show() if true (default is True). Sometimes
we want to draw multiple things before calling plt.show()
:type show: bool
:return: None
SAMPLE CODE
rf = RandomForestRegressor(n_estimators=100, n_jobs=-1, oob_score=True)
X_train, y_train = ..., ...
rf.fit(X_train, y_train)
imp = importances(rf, X_test, y_test)
plot_importances(imp)
"""
I = df_importances
# this is backwards but seems to undo weird reversed order in barh()
I = I.sort_values('Importance', ascending=True)
if figsize:
fig = plt.figure(figsize=figsize)
elif scalefig:
fig = plt.figure()
w, h = fig.get_size_inches()
fig.set_size_inches(w * scalefig[0], h * scalefig[1], forward=True)
else:
fig = plt.figure()
ax = plt.gca()
labels = []
for col in I.index:
if isinstance(col, list):
labels.append('\n'.join(col))
else:
labels.append(col)
for tick in ax.get_xticklabels():
tick.set_size(label_fontsize)
for tick in ax.get_yticklabels():
tick.set_size(label_fontsize)
ax.barh(np.arange(len(I.index)), I.Importance, height=0.6, tick_label=labels)
# rotate x-ticks
if xrot is not None:
plt.xticks(rotation=xrot)
# xticks freq
xticks = ax.get_xticks()
nticks = len(xticks)
new_ticks = xticks[np.arange(0, nticks, step=tickstep)]
ax.set_xticks(new_ticks)
if save:
plt.savefig(save, bbox_inches="tight", pad_inches=0.03)
if show:
plt.show()
def oob_dependences(rf, X_train, n_samples=5000):
"""
Given a random forest model, rf, and training observation independent
variables in X_train (a dataframe), compute the OOB R^2 score using each var
as a dependent variable. We retrain rf for each var. Only numeric columns are considered.
By default, sample up to 5000 observations to compute feature dependencies.
:return: Return a DataFrame with Feature/Dependence values for each variable. Feature is the dataframe index.
"""
numcols = [col for col in X_train if is_numeric_dtype(X_train[col])]
X_train = sample_rows(X_train, n_samples)
df_dep = pd.DataFrame(columns=['Feature','Dependence'])
df_dep = df_dep.set_index('Feature')
for col in numcols:
X, y = X_train.drop(col, axis=1), X_train[col]
rf.fit(X, y)
df_dep.loc[col] = rf.oob_score_
df_dep = df_dep.sort_values('Dependence', ascending=False)
return df_dep
def feature_dependence_matrix(rf, X_train, n_samples=5000):
"""
Given training observation independent variables in X_train (a dataframe),
compute the feature importance using each var as a dependent variable.
We retrain a random forest for each var as target using the others as
independent vars. Only numeric columns are considered.
By default, sample up to 5000 observations to compute feature dependencies.
:return: a non-symmetric data frame with the dependence matrix where each row is the importance of each var to the row's var used as a model target.
"""
numcols = [col for col in X_train if is_numeric_dtype(X_train[col])]
X_train = sample_rows(X_train, n_samples)
df_dep = pd.DataFrame(index=X_train.columns, columns=['Dependence']+X_train.columns.tolist())
for i in range(len(numcols)):
col = numcols[i]
X, y = X_train.drop(col, axis=1), X_train[col]
rf.fit(X,y)
#imp = rf.feature_importances_
imp = permutation_importances_raw(rf, X, y, oob_regression_r2_score, n_samples)
imp = np.insert(imp, i, 1.0)
df_dep.iloc[i] = np.insert(imp, 0, rf.oob_score_) # add overall dependence
return df_dep
def feature_corr_matrix(df):
"""
Return the Spearman's rank-order correlation between all pairs
of features as a matrix with feature names as index and column names.
The diagonal will be all 1.0 as features are self correlated.
Spearman's correlation is the same thing as converting two variables
to rank values and then running a standard Pearson's correlation
on those ranked variables. Spearman's is nonparametric and does not
assume a linear relationship between the variables; it looks for
monotonic relationships.
:param df_train: dataframe containing features as columns, and
without the target variable.
:return: a data frame with the correlation matrix
"""
corr = np.round(spearmanr(df).correlation, 4)
df_corr = pd.DataFrame(data=corr, index=df.columns, columns=df.columns)
return df_corr
def plot_corr_heatmap(df,
threshold=0.6,
cmap=None,
figsize=None,
value_fontsize=12, label_fontsize=14,
xrot=80,
save=None,
show=True):
"""
Display the feature spearman's correlation matrix as a heatmap with
any abs(value)>threshold appearing with background color.
Spearman's correlation is the same thing as converting two variables
to rank values and then running a standard Pearson's correlation
on those ranked variables. Spearman's is nonparametric and does not
assume a linear relationship between the variables; it looks for
monotonic relationships.
SAMPLE CODE
from rfpimp import plot_corr_heatmap
plot_corr_heatmap(df_train, save='/tmp/corrheatmap.svg',
figsize=(7,5), label_fontsize=13, value_fontsize=11)
"""
corr = np.round(spearmanr(df).correlation, 4)
filtered = copy(corr)
filtered = np.abs(filtered) # work with abs but display negatives later
mask = np.ones_like(corr)
filtered[np.tril_indices_from(mask)] = -9999
if not cmap:
cw = plt.get_cmap('coolwarm')
cmap = ListedColormap([cw(x) for x in np.arange(.6, .85, 0.01)])
elif isinstance(cmap, str):
cmap = plt.get_cmap(cmap)
cm = copy(cmap)
cm.set_under(color='white')
if figsize:
plt.figure(figsize=figsize)
plt.imshow(filtered, cmap=cm, vmin=threshold, vmax=1, aspect='equal')
width, height = filtered.shape
for x in range(width):
for y in range(height):
if x < y:
plt.annotate(str(np.round(corr[x, y], 2)), xy=(y, x),
horizontalalignment='center',
verticalalignment='center',
fontsize=value_fontsize)
plt.colorbar()
plt.xticks(range(width), df.columns, rotation=xrot, horizontalalignment='right',
fontsize=label_fontsize)
plt.yticks(range(width), df.columns, verticalalignment='center',
fontsize=label_fontsize)
if save:
plt.savefig(save, bbox_inches="tight", pad_inches=0.03)
if show:
plt.show()
def jeremy_trick_RF_sample_size(n):
# Jeremy's trick; hmm.. this won't work as a separate function?
# def batch_size_for_node_splitting(rs, n_samples):
# forest.check_random_state(rs).randint(0, n_samples, 20000)
# forest._generate_sample_indices = batch_size_for_node_splitting
forest._generate_sample_indices = \
(lambda rs, n_samples: forest.check_random_state(rs).randint(0, n_samples, n))
def jeremy_trick_reset_RF_sample_size():
forest._generate_sample_indices = (lambda rs, n_samples:
forest.check_random_state(rs).randint(0, n_samples, n_samples))
