Numerical features are the easiest data type to feed a model directly — but "already a number" doesn't mean "already in the right form." Binning, transforming and scaling numeric features correctly often matters as much as any encoding decision for categorical data.
Common Numerical Feature Engineering Techniques
| Technique | What It Does | When to Use |
|---|---|---|
| Binning / discretization | Converts a continuous value into ranges (e.g. age → "18-25", "26-35", ...) | When the relationship to the target is non-linear but roughly step-wise |
| Log / power transforms | Compresses right-skewed distributions closer to symmetric | Income, prices, counts — see Feature Transformation |
| Scaling | Rescales a feature's range | Distance- and gradient-based models |
| Ratios and rates | Combines two numeric features into one more meaningful signal | "income / household_size" often predicts more than either alone |
Binning — A Worked Example
import pandas as pd
df = pd.DataFrame({"age": [17, 22, 29, 35, 44, 58, 65]})
df["age_group"] = pd.cut(
df["age"],
bins=[0, 18, 30, 45, 60, 100],
labels=["under_18", "18_30", "30_45", "45_60", "60_plus"]
)
print(df)Binning trades away some precision (35 and 44 both fall in "30_45") in exchange for letting a linear model capture a non-linear age effect — e.g. if risk is elevated specifically for both very young and very old customers, a single linear "age" coefficient can't represent that U-shape, but a set of age-group dummy variables can.
Ratios — Often More Predictive Than Either Raw Feature
df["debt_to_income"] = df["total_debt"] / df["annual_income"]
df["price_per_sqft"] = df["price"] / df["size_sqft"]A ratio like debt-to-income is frequently far more predictive for credit risk than either "debt" or "income" alone — it directly encodes the relationship a domain expert would actually reason about, instead of leaving the model to rediscover that relationship on its own from two separate columns.
Practical Use Cases
- Converting a continuous feature with a non-linear target relationship into bins a linear model can use
- Building domain-informed ratios (price per unit, rate per time period) that carry more signal than raw values
- Reducing the influence of extreme values through binning, as an alternative to outlier removal
Common Mistakes
- Binning a feature that already has a genuinely linear relationship with the target — this throws away precision for no benefit.
- Choosing bin edges arbitrarily instead of based on domain knowledge or the data's actual distribution (e.g. quantile-based bins via
pd.qcut). - Creating a ratio with a denominator that can be zero, without handling the resulting division error.
Interview Relevance
Q: "When would you bin a numeric feature instead of leaving it continuous?" When you suspect a non-linear, step-wise relationship with the target that a linear model can't otherwise capture, or when you specifically want to reduce sensitivity to noise/outliers at the cost of some precision.
Practice Question
You have "num_dependents" and "annual_income" columns for a loan approval model. Propose one ratio feature that would likely help a linear model more than the two raw columns alone.
