Data

X, y = make_blobs(n_samples=100, centers=3, random_state=12, cluster_std=5) # make synthetic data
X_train_toy, X_test_toy, y_train_toy, y_test_toy = train_test_split(
    X, y, random_state=5, test_size=0.4) # split it into training and test sets
# Visualize the training data
plt.scatter(X_train_toy[:, 0], X_train_toy[:, 1], label="Training set", s=60)
plt.scatter(
    X_test_toy[:, 0], X_test_toy[:, 1], color=mglearn.cm2(1), label="Test set", s=60
)
plt.legend(loc="upper right")

Bad methodology 1

• What’s wrong with scaling data separately?

scaler = StandardScaler() # Creating a scalert object 
scaler.fit(X_train_toy) # Calling fit on the training data 
train_scaled = scaler.transform(
    X_train_toy
)  # Transforming the training data using the scaler fit on training data

scaler = StandardScaler()  # Creating a separate object for scaling test data
scaler.fit(X_test_toy)  # Calling fit on the test data
test_scaled = scaler.transform(
    X_test_toy
)  # Transforming the test data using the scaler fit on test data

knn = KNeighborsClassifier()
knn.fit(train_scaled, y_train_toy)
print(f"Training score: {knn.score(train_scaled, y_train_toy):.2f}")
print(f"Test score: {knn.score(test_scaled, y_test_toy):.2f}") # misleading scores

Training score: 0.63
Test score: 0.60

Scaling train and test data separately

Bad methodology 2

• What’s wrong with scaling the data together

# join the train and test sets back together
XX = np.vstack((X_train_toy, X_test_toy))

scaler = StandardScaler()
scaler.fit(XX)
XX_scaled = scaler.transform(XX)

XX_train = XX_scaled[:X_train_toy.shape[0]]
XX_test = XX_scaled[X_train_toy.shape[0]:]

knn = KNeighborsClassifier()
knn.fit(XX_train, y_train_toy)
print(f"Training score: {knn.score(XX_train, y_train_toy):.2f}")  # Misleading score
print(f"Test score: {knn.score(XX_test, y_test_toy):.2f}")  # Misleading score

Training score: 0.63
Test score: 0.55

Bad methodology 3

• What’s wrong here?

knn = KNeighborsClassifier()

scaler = StandardScaler()
scaler.fit(X_train_toy)
X_train_scaled = scaler.transform(X_train_toy)
X_test_scaled = scaler.transform(X_test_toy)
cross_val_score(knn, X_train_scaled, y_train_toy)

array([0.25      , 0.5       , 0.58333333, 0.58333333, 0.41666667])

Recap: sklearn Pipelines

• Pipeline is a way to chain multiple steps (e.g., preprocessing + model fitting) into a single workflow.
• Simplify the code and improves readability.
• Reduce the risk of data leakage by ensuring proper transformation of the training and test sets.
• Automatically apply transformations in sequence.
• Example:
  • Chaining a StandardScaler with a KNeighborsClassifier model.

from sklearn.pipeline import make_pipeline

pipe_knn = make_pipeline(StandardScaler(), KNeighborsClassifier())

# Correct way to do cross validation without breaking the golden rule. 
cross_val_score(pipe_knn, X_train_toy, y_train_toy) 

array([0.25      , 0.5       , 0.5       , 0.58333333, 0.41666667])

sklearn’s ColumnTransformer

• Use ColumnTransformer to build all our transformations together into one object

• Use a column transformer with sklearn pipelines.

(iClicker) Exercise 6.1

iClicker cloud join link: https://join.iclicker.com/VYFJ

Select all of the following statements which are TRUE.

• 1. You could carry out cross-validation by passing a ColumnTransformer object to cross_validate.
• 2. After applying column transformer, the order of the columns in the transformed data has to be the same as the order of the columns in the original data.
• 3. After applying a column transformer, the transformed data is always going to be of different shape than the original data.
• 4. When you call fit_transform on a ColumnTransformer object, you get a numpy ndarray.

Remarks on preprocessing

• There is no one-size-fits-all solution in data preprocessing, and decisions often involve a degree of subjectivity.
  • Exploratory data analysis and domain knowledge inform these decisions
• Always consider the specific goals of your project when deciding how to encode features.

Alternative methods for scaling

• StandardScaler
  • Good choice when the column follows a normal distribution or a distribution somewhat like a normal distribution.
• MinMaxScaler: Transform each feature to a desired range. Appropriate when
  • Good choice for features such as human age, where there is a fixed range of values and the feature is uniformly distributed across the range
• Normalizer: Works on rows rather than columns. Normalize examples individually to unit norm.
  • Good choice for frequency-type data
• Log scaling
  • Good choice for features such as ratings per movies (power law distribution; a few movies have lots of ratings but most movies have very few ratings)
• …

Ordinal encoding vs. One-hot encoding

• Ordinal Encoding: Encodes categorical features as an integer array.
• One-hot Encoding: Creates binary columns for each category’s presence.
• Sometimes how we encode a specific feature depends upon the context.

Ordinal encoding vs. One-hot encoding

• Consider weather feature and its four categories: Sunny (☀️), Cloudy (🌥️), Rainy (⛈️), Snowy (❄️)
• Which encoding would you use in each of the following scenarios?
  • Predicting traffic volume
  • Predicting severity of weather-related road incidents

Ordinal encoding vs. One-hot encoding

• Consider weather feature and its four categories: Sunny (☀️), Cloudy (🌥️), Rainy (⛈️), Snowy (❄️)
• Predicting traffic volume: Using one-hot encoding would make sense here because the impact of different weather conditions on traffic volume does not necessarily follow a clear order and different weather conditions could have very distinct effects.
• Predicting severity of weather-related road incidents: An ordinal encoding might be more appropriate if you define your weather categories from least to most severe as this could correlate directly with the likelihood or severity of incidents.

handle_unknown = "ignore" of OneHotEncoder

• Use handle_unknown='ignore' with OneHotEncoder to safely ignore unseen categories during transform.
• In each of the following scenarios, identify whether it’s a reasonable strategy or not.
  • Example 1: Suppose you are building a model to predict customer behavior (e.g., purchase likelihood) based on features like location, device_type, and product_category. During training, you have observed a set of categories for product_category, but in the future, new product categories might be added.
  • Example 2: You’re building a model to predict disease diagnosis based on symptoms, where each symptom is categorized (e.g., fever, headache, nausea).

handle_unknown = "ignore" of OneHotEncoder

• Reasonable use: When unseen categories are less likely to impact the model’s prediction accuracy (e.g., product categories in e-commerce), and you prefer to avoid breaking the model.
• Not-so-reasonable use: When unseen categories could provide critical new information that could significantly alter predictions (e.g., in medical diagnostics), ignoring them could result in a poor or dangerous outcome.

drop="if_binary" argument of OneHotEncoder

• drop=‘if_binary’ argument in OneHotEncoder:
• Reduces redundancy by dropping one of the columns if the feature is binary.

Categorical variables with too many categories

• Strategies for categorical variables with too many categories:
  • Dimensionality reduction techniques
  • Bucketing categories into ‘others’
  • Clustering or grouping categories manually
  • Only considering top-N categories
  • …

Dealing with text features

• Preprocessing text to fit into machine learning models using text vectorization.
• Bag of words representation

sklearn CountVectorizer

• Use scikit-learn’s CountVectorizer to encode text data
• CountVectorizer: Transforms text into a matrix of token counts
• Important parameters:
  • max_features: Control the number of features used in the model
  • max_df, min_df: Control document frequency thresholds
  • ngram_range: Defines the range of n-grams to be extracted
  • stop_words: Enables the removal of common words that are typically uninformative in most applications, such as “and”, “the”, etc.

Incorporating text features in a machine learning pipeline

from sklearn.feature_extraction.text import CountVectorizer
from sklearn.svm import SVC
from sklearn.pipeline import make_pipeline

text_pipeline = make_pipeline(
    CountVectorizer(),
    SVC()
)

(iClicker) Exercise 6.2

iClicker cloud join link: https://join.iclicker.com/VYFJ

Select all of the following statements which are TRUE.

• 1. handle_unknown="ignore" would treat all unknown categories equally.
• 2. As you increase the value for max_features hyperparameter of CountVectorizer the training score is likely to go up.
• 3. Suppose you are encoding text data using CountVectorizer. If you encounter a word in the validation or the test split that’s not available in the training data, we’ll get an error.
• 4. In the code below, inside cross_validate, each fold might have slightly different number of features (columns) in the fold.

pipe = (CountVectorizer(), SVC())
cross_validate(pipe, X_train, y_train)