Machine Learning
Contents
6.5. Machine Learning#
6.5.1. causalimpact: Find Causal Relation of an Event and a Variable in Python#
!pip install pycausalimpact
When working with time series data, you might want to determine whether an event has an impact on some response variable or not. For example, if your company creates an advertisement, you might want to track whether the advertisement results in an increase in sales or not.
That is when causalimpact comes in handy. causalimpact analyses the differences between expected and observed time series data. With causalimpact, you can infer the expected effect of an intervention in 3 lines of code.
import numpy as np
import pandas as pd
from statsmodels.tsa.arima_process import ArmaProcess
import causalimpact
from causalimpact import CausalImpact
# Generate random sample
np.random.seed(0)
ar = np.r_[1, 0.9]
ma = np.array([1])
arma_process = ArmaProcess(ar, ma)
X = 50 + arma_process.generate_sample(nsample=1000)
y = 1.6 * X + np.random.normal(size=1000)
# There is a change starting from index 800
y[800:] += 10
data = pd.DataFrame({"y": y, "X": X}, columns=["y", "X"])
pre_period = [0, 799]
post_period = [800, 999]
ci = CausalImpact(data, pre_period, post_period)
print(ci.summary())
print(ci.summary(output="report"))
ci.plot()
Posterior Inference {Causal Impact}
Average Cumulative
Actual 90.03 18006.16
Prediction (s.d.) 79.97 (0.3) 15994.43 (60.93)
95% CI [79.36, 80.55] [15871.15, 16110.0]
Absolute effect (s.d.) 10.06 (0.3) 2011.72 (60.93)
95% CI [9.48, 10.68] [1896.16, 2135.01]
Relative effect (s.d.) 12.58% (0.38%) 12.58% (0.38%)
95% CI [11.86%, 13.35%] [11.86%, 13.35%]
Posterior tail-area probability p: 0.0
Posterior prob. of a causal effect: 100.0%
For more details run the command: print(impact.summary('report'))
Analysis report {CausalImpact}
During the post-intervention period, the response variable had
an average value of approx. 90.03. By contrast, in the absence of an
intervention, we would have expected an average response of 79.97.
The 95% interval of this counterfactual prediction is [79.36, 80.55].
Subtracting this prediction from the observed response yields
an estimate of the causal effect the intervention had on the
response variable. This effect is 10.06 with a 95% interval of
[9.48, 10.68]. For a discussion of the significance of this effect,
see below.
Summing up the individual data points during the post-intervention
period (which can only sometimes be meaningfully interpreted), the
response variable had an overall value of 18006.16.
By contrast, had the intervention not taken place, we would have expected
a sum of 15994.43. The 95% interval of this prediction is [15871.15, 16110.0].
The above results are given in terms of absolute numbers. In relative
terms, the response variable showed an increase of +12.58%. The 95%
interval of this percentage is [11.86%, 13.35%].
This means that the positive effect observed during the intervention
period is statistically significant and unlikely to be due to random
fluctuations. It should be noted, however, that the question of whether
this increase also bears substantive significance can only be answered
by comparing the absolute effect (10.06) to the original goal
of the underlying intervention.
The probability of obtaining this effect by chance is very small
(Bayesian one-sided tail-area probability p = 0.0).
This means the causal effect can be considered statistically
significant.
6.5.2. Create a Readable Machine Learning Pipeline in One Line of Code#
If you want to create a readable machine learning pipeline in a single line of code, try the make_pipeline function in scikit-learn.
make_pipeline is especially useful when working with complex pipelines that involve many different transformers and estimators.
from sklearn.datasets import make_classification
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
X, y = make_classification(random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)
# Create a pipeline that scales the data and fits a logistic regression model
pipeline = make_pipeline(StandardScaler(), LogisticRegression())
# Fit the pipeline to the training data
pipeline.fit(X_train, y_train)
# Evaluate the pipeline on the test data
pipeline.score(X_test, y_test)
0.96
6.5.3. Pipeline + GridSearchCV: Prevent Data Leakage when Scaling the Data#
Scaling the data before using GridSearchCV can lead to data leakage since the scaling tells some information about the entire data. To prevent this, assemble both the scaler and machine learning models in a pipeline then use it as the estimator for GridSearchCV.
from sklearn.model_selection import train_test_split, GridSearchCV
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import make_pipeline
from sklearn.svm import SVC
from sklearn.datasets import load_iris
# load data
df = load_iris()
X = df.data
y = df.target
# split data into train and test
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=0)
# Create a pipeline variable
make_pipe = make_pipeline(StandardScaler(), SVC())
# Defining parameters grid
grid_params = {"svc__C": [0.1, 1, 10, 100, 1000], "svc__gamma": [0.1, 1, 10, 100]}
# hypertuning
grid = GridSearchCV(make_pipe, grid_params, cv=5)
grid.fit(X_train, y_train)
# predict
y_pred = grid.predict(X_test)
The estimator is now the entire pipeline instead of just the machine learning model.
6.5.4. squared=False: Get RMSE from Sklearn’s mean_squared_error method#
If you want to get the root mean squared error using sklearn, pass squared=False to sklearn’s mean_squared_error method.
from sklearn.metrics import mean_squared_error
y_actual = [1, 2, 3]
y_predicted = [1.5, 2.5, 3.5]
rmse = mean_squared_error(y_actual, y_predicted, squared=False)
rmse
0.5
6.5.5. modelkit: Build Production ML Systems in Python#
!pip install modelkit textblob
If you want your ML models to be fast, type-safe, testable, and fast to deploy to production, try modelkit. modelkit allows you to incorporate all of these features into your model in several lines of code.
from modelkit import ModelLibrary, Model
from textblob import TextBlob, WordList
# import nltk
# nltk.download('brown')
# nltk.download('punkt')
To define a modelkit Model, you need to:
create class inheriting from
modelkit.Modelimplement a
_predictmethod
class NounPhraseExtractor(Model):
# Give model a name
CONFIGURATIONS = {"noun_phrase_extractor": {}}
def _predict(self, text):
blob = TextBlob(text)
return blob.noun_phrases
You can now instantiate and use the model:
noun_extractor = NounPhraseExtractor()
noun_extractor("What are your learning strategies?")
2021-11-05 09:55.55 [debug ] Model loaded memory=0 Bytes memory_bytes=0 model_name=None time=0 microseconds time_s=4.232699939166196e-05
WordList(['learning strategies'])
You can also create test cases for your model and make sure all test cases are passed.
class NounPhraseExtractor(Model):
# Give model a name
CONFIGURATIONS = {"noun_phrase_extractor": {}}
TEST_CASES = [
{"item": "There is a red apple on the tree", "result": WordList(["red apple"])}
]
def _predict(self, text):
blob = TextBlob(text)
return blob.noun_phrases
noun_extractor = NounPhraseExtractor()
noun_extractor.test()
2021-11-05 09:55.58 [debug ] Model loaded memory=0 Bytes memory_bytes=0 model_name=None time=0 microseconds time_s=4.3191997974645346e-05
TEST 1: SUCCESS
modelkit also allows you to organize a group of models using ModelLibrary.
class SentimentAnalyzer(Model):
# Give model a name
CONFIGURATIONS = {"sentiment_analyzer": {}}
def _predict(self, text):
blob = TextBlob(text)
return blob.sentiment
nlp_models = ModelLibrary(models=[NounPhraseExtractor, SentimentAnalyzer])
2021-11-05 09:50.13 [info ] Instantiating AssetsManager lazy_loading=False
2021-11-05 09:50.13 [info ] No remote storage provider configured
2021-11-05 09:50.13 [debug ] Resolving asset for Model model_name=sentiment_analyzer
2021-11-05 09:50.13 [debug ] Loading model model_name=sentiment_analyzer
2021-11-05 09:50.13 [debug ] Instantiating Model object model_name=sentiment_analyzer
2021-11-05 09:50.13 [debug ] Model loaded memory=0 Bytes memory_bytes=0 model_name=sentiment_analyzer time=0 microseconds time_s=3.988200114690699e-05
2021-11-05 09:50.13 [debug ] Done loading Model model_name=sentiment_analyzer
2021-11-05 09:50.13 [info ] Model and dependencies loaded memory=0 Bytes memory_bytes=0 name=sentiment_analyzer time=0 microseconds time_s=0.00894871700074873
2021-11-05 09:50.13 [debug ] Resolving asset for Model model_name=noun_phrase_extractor
2021-11-05 09:50.13 [debug ] Loading model model_name=noun_phrase_extractor
2021-11-05 09:50.13 [debug ] Instantiating Model object model_name=noun_phrase_extractor
2021-11-05 09:50.13 [debug ] Model loaded memory=0 Bytes memory_bytes=0 model_name=noun_phrase_extractor time=0 microseconds time_s=2.751099964370951e-05
2021-11-05 09:50.13 [debug ] Done loading Model model_name=noun_phrase_extractor
2021-11-05 09:50.13 [info ] Model and dependencies loaded memory=0 Bytes memory_bytes=0 name=noun_phrase_extractor time=0 microseconds time_s=0.006440052002290031
Get and use the models from nlp_models.
noun_extractor = model_collections.get("noun_phrase_extractor")
noun_extractor("What are your learning strategies?")
WordList(['learning strategies'])
sentiment_analyzer = model_collections.get("sentiment_analyzer")
sentiment_analyzer("Today is a beautiful day!")
Sentiment(polarity=1.0, subjectivity=1.0)
6.5.6. Decompose high dimensional data into two or three dimensions#
!pip install yellowbrick
If you want to decompose high dimensional data into two or three dimensions to visualize it, what should you do? A common technique is PCA. Even though PCA is useful, it can be complicated to create a PCA plot.
Lucikily, Yellowbrick allows you visualize PCA in a few lines of code
from yellowbrick.datasets import load_credit
from yellowbrick.features import PCA
X, y = load_credit()
classes = ["account in defaut", "current with bills"]
visualizer = PCA(scale=True, classes=classes)
visualizer.fit_transform(X, y)
visualizer.show()
<AxesSubplot:title={'center':'Principal Component Plot'}, xlabel='$PC_1$', ylabel='$PC_2$'>
6.5.7. Visualize Feature Importances with Yellowbrick#
!pip install yellowbrick
Having more features is not always equivalent to a better model. The more features a model has, the more sensitive the model is to errors due to variance. Thus, we want to select the minimum required features to produce a valid model.
A common approach to eliminate features is to eliminate the ones that are the least important to the model. Then we re-evaluate if the model actually performs better during cross-validation.
Yellowbrick’s FeatureImportances is ideal for this task since it helps us to visualize the relative importance of the features for the model.
from sklearn.tree import DecisionTreeClassifier
from yellowbrick.datasets import load_occupancy
from yellowbrick.model_selection import FeatureImportances
X, y = load_occupancy()
model = DecisionTreeClassifier()
viz = FeatureImportances(model)
viz.fit(X, y)
viz.show();
From the plot above, it seems like the light is the most important feature to DecisionTreeClassifier, followed by CO2, temperature.
6.5.8. Validation Curve: Determine if an Estimator Is Underfitting Over Overfitting#
!pip install yellowbrick
To find the hyperparameter where the estimator is neither underfitting nor overfitting, use Yellowbrick’s validation curve.
from yellowbrick.datasets.loaders import load_occupancy
from yellowbrick.model_selection import validation_curve
from sklearn.tree import DecisionTreeClassifier
import numpy as np
# Load data
X, y = load_occupancy()
In the code below, we choose the range of max_depth to be from 1 to 11.
viz = validation_curve(
DecisionTreeClassifier(), X, y, param_name="max_depth",
param_range=np.arange(1, 11), cv=10, scoring="f1",
)
As we can see from the plot above, although max_depth > 2 has a higher training score but a lower cross-validation score. This indicates that the model is overfitting.
Thus, the sweet spot will be where the cross-validation score neither increases nor decreases, which is 2.
6.5.9. Mlxtend: Plot Decision Regions of Your ML Classifiers#
!pip install mlxtend
How does your machine learning classifier decide which class a sample belongs to? Plotting a decision region can give you some insights into your ML classifier’s decision.
An easy way to plot decision regions is to use mlxtend’s plot_decision_regions.
import matplotlib.pyplot as plt
import matplotlib.gridspec as gridspec
import itertools
from sklearn.linear_model import LogisticRegression
from sklearn.svm import SVC
from sklearn.ensemble import RandomForestClassifier
from mlxtend.classifier import EnsembleVoteClassifier
from mlxtend.data import iris_data
from mlxtend.plotting import plot_decision_regions
# Initializing Classifiers
clf1 = LogisticRegression(random_state=0)
clf2 = RandomForestClassifier(random_state=0)
clf3 = SVC(random_state=0, probability=True)
eclf = EnsembleVoteClassifier(clfs=[clf1, clf2, clf3], weights=[2, 1, 1], voting='soft')
# Loading some example data
X, y = iris_data()
X = X[:,[0, 2]]
# Plotting Decision Regions
gs = gridspec.GridSpec(2, 2)
fig = plt.figure(figsize=(10, 8))
for clf, lab, grd in zip([clf1, clf2, clf3, eclf],
['Logistic Regression', 'Random Forest', 'RBF kernel SVM', 'Ensemble'],
itertools.product([0, 1], repeat=2)):
clf.fit(X, y)
ax = plt.subplot(gs[grd[0], grd[1]])
fig = plot_decision_regions(X=X, y=y, clf=clf, legend=2)
plt.title(lab)
plt.show()
Mlxtend (machine learning extensions) is a Python library of useful tools for the day-to-day data science tasks.
Find other useful functionalities of Mlxtend here.
6.5.10. Deepchecks + Weights & Biases: Test and Track Your ML Model and Data#
!pip install -U deepchecks wandb scikit-learn
Weight and Biases is a tool to track and monitor your ML experiments. deepchecks is a tool that allows you to create test suites for your ML models & data with ease.
The checks in a suite includes:
model performance
data integrity
distribution mismatches and more.
Now you can track deepchecks suite’s results with Weights & Biases.
To learn how to use this integration, start with loading a dataset and splitting it into train and test set.
import pandas as pd
import numpy as np
from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
from deepchecks import Dataset
import warnings
warnings.simplefilter(action='ignore', category=FutureWarning)
# Load Data
X, y = load_iris(as_frame=True, return_X_y=True)
label_col = 'target'
df = pd.concat([X, y.to_frame(name=label_col)], axis=1)
df_train, df_test = train_test_split(df, stratify=df[label_col], random_state=0)
Next, build a ML model using the training data.
rf_clf = RandomForestClassifier()
rf_clf.fit(df_train.drop(label_col, axis=1), df_train[label_col])
RandomForestClassifier()
Create deepchecks’ Dataset objects with train and test set.
ds_train = Dataset(df_train, label=label_col, cat_features=[])
ds_test = Dataset(df_test, label=label_col, cat_features=[])
Create a test suite using our ML model and datasets and run it.
from deepchecks.suites import full_suite
suite = full_suite()
suite_result = suite.run(ds_train, ds_test, rf_clf)
Export all results to Weights & Biases:
import wandb
wandb.login()
wandb: Currently logged in as: khuyentran1401 (use `wandb login --relogin` to force relogin)
True
suite_result.to_wandb()
/home/khuyen/book/book/Chapter5/wandb/run-20220314_094658-1yf63l3g
Synced 4 W&B file(s), 109 media file(s), 90 artifact file(s) and 0 other file(s)
./wandb/run-20220314_094658-1yf63l3g/logsAnd this is how the test suite will look like in W&B.
6.5.11. imbalanced-learn: Deal with an Imbalanced Dataset#
!pip install imbalanced-learn==0.10.0 mlxtend==0.21.0
A dataset is imbalanced when the number of samples in one class is much more than the other classes. When training a sensitive classifier using an imbalanced dataset, it will work well on the majority class but work poorly on the minority class.
To deal with an imbalanced dataset, we can use imbalanced-learn to generate new samples in the classes which are under-represented.
In the example below, we use the RandomOverSampler class from imbalanced-learn to generate new samples by randomly sampling with replacement the current available samples.
# Libraries for plotting
import matplotlib.pyplot as plt
import seaborn as sns
from mlxtend.plotting import plot_decision_regions
import matplotlib.gridspec as gridspec
# Libraries for machine learning
from sklearn.datasets import make_classification
from sklearn.svm import LinearSVC
import warnings
warnings.simplefilter("ignore", UserWarning)
X, y = make_classification(
n_samples=5000,
n_features=2,
n_informative=2,
n_redundant=0,
n_repeated=0,
n_classes=4,
n_clusters_per_class=1,
weights=[0.01, 0.04, 0.5, 0.90],
class_sep=0.8,
random_state=0,
)
from imblearn.over_sampling import RandomOverSampler
ros = RandomOverSampler(random_state=1)
X_resampled, y_resampled = ros.fit_resample(X, y)
# Plotting Decision Regions
fig, (ax0, ax1) = plt.subplots(nrows=2, ncols=1, sharey=True, figsize=(6, 10))
for Xi, yi, ax, title in zip(
[X, X_resampled],
[y, y_resampled],
[ax0, ax1],
["Without resampling", "Using RandomOverSampler"],
):
clf = LinearSVC()
clf.fit(Xi, yi)
fig = plot_decision_regions(X=Xi, y=yi, clf=clf, legend=2, ax=ax)
plt.title(title)
6.5.12. Estimate Prediction Intervals in Scikit-Learn Models with MAPIE#
!pip install mapie
To get estimated prediction intervals for predictions made by a scikit-learn model, use MAPIE.
In the code below, we use MapieRegressor to estimate prediction intervals for a scikit-learn regressor.
from mapie.regression import MapieRegressor
from sklearn.linear_model import LinearRegression
from sklearn.datasets import make_regression
# Create data
X, y = make_regression(n_samples=200, n_features=1, noise=50, random_state=0)
# Train and predict
alpha = [0.05, 0.32]
mapie = MapieRegressor(LinearRegression())
mapie.fit(X, y)
y_pred, y_pis = mapie.predict(X, alpha=alpha)
# compute the coverage of the prediction intervals
from mapie.metrics import regression_coverage_score
coverage_scores = [
regression_coverage_score(y, y_pis[:, 0, i], y_pis[:, 1, i])
for i, _ in enumerate(alpha)
]
# Plot the estimated prediction intervals
from matplotlib import pyplot as plt
import numpy as np
plt.figure(figsize=(11, 7))
plt.xlabel("x")
plt.ylabel("y")
plt.scatter(X, y, alpha=0.3)
plt.plot(X, y_pred, color="C1")
order = np.argsort(X[:, 0])
plt.plot(X[order], y_pis[order][:, 0, 1], color="C1", ls="--")
plt.plot(X[order], y_pis[order][:, 1, 1], color="C1", ls="--")
plt.fill_between(
X[order].ravel(),
y_pis[order][:, 0, 0].ravel(),
y_pis[order][:, 1, 0].ravel(),
alpha=0.2,
)
plt.title(
f"Target coverages for "
f"alpha={alpha[0]:.2f}: ({1-alpha[0]:.3f}, {coverage_scores[0]:.3f})\n"
f"Target coverages for "
f"alpha={alpha[1]:.2f}: ({1-alpha[1]:.3f}, {coverage_scores[1]:.3f})"
)
plt.show()
