Summary - Feature engineering

In this group you learned feature selection, extraction, and transformation.

Summary: feature engineering

In this group:

Features are the only information the model sees; a good feature relates to the target and is known at prediction time.
Simple selection is done with a variance threshold and correlation.
Extraction from dates and text: month, day, word counts.
Categorical features: one-hot (unordered), ordinal (ordered), handle_unknown="ignore".
Numeric transformations: log1p, binning, polynomials.

Every transformation must be learned on training data and applied unchanged to test data; a pipeline is the safest way.

Exercise Features from a date

Build a list containing the month and the weekend flag (1 if weekday() >= 5, else 0) for each date and print it as an np.array.

Show example solution

import numpy as np
from datetime import date

tarihler = [date(2024, 1, 6), date(2024, 1, 8), date(2024, 7, 14)]

ozet = [[t.month, int(t.weekday() >= 5)] for t in tarihler]
print(np.array(ozet))

Test yourself

Which carries a leakage risk for a credit model?
- Income at application time
- The applicant's age
- The amount requested
- The number of late payments computed after the loan closed
Show answer

The number of late payments computed after the loan closed — This information doesn't exist at prediction time; it is derived from the target itself.
Compress the skewed value with a logarithm.
```
x_log = np.____(gelir)
```
Show answer
```
x_log = np.log1p(gelir)
```
Match the tools.

VarianceThreshold

Drops constant columns
CountVectorizer

Counts words
OneHotEncoder

Category columns
PolynomialFeatures

Squares and products

Next lesson: What is model selection?

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies