An online store is predicting future book sales by using a linear regression model that is based on past sales data. The data includes duration, a numerical feature that represents the number of…

Question

Options

• AOne-hot encoding
• BCartesian product transformation
• CQuantile binning
• DNormalization

• A
  9% (5)
• B
  5% (3)
• C
  84% (48)
• D
  2% (1)

Explanation

When a numerical feature has a skewed, non-linear relationship with the target variable, quantile binning (C) is the appropriate transformation. Quantile binning divides the feature values into equal-frequency buckets (e.g., quartiles or deciles), effectively converting the continuous skewed variable into an ordinal or one-hot-encoded categorical feature. This removes the assumption of linearity between 'duration' and 'sales' and allows the linear regression model to capture non-linear patterns across different ranges of the feature. One-hot encoding (A) is for categorical features, not continuous numerical ones. Cartesian product transformation (B) creates interaction terms between features but does not address skewness in a single feature. Normalization (D) rescales the range of values but preserves the underlying non-linear distribution and does not address the skewed relationship with the target.

No community discussion yet for this question.