Missing as its own category

Treating missing as its own category means replacing the gaps in a text column with a label such as Missing, so the absence becomes a value the model can learn from instead of a hole that has to be patched.

Last updated: 12 Oct, 2026 · pandas 3.0

Frequent category imputation filled a column that was 47.26 percent missing with its mode and the median price of the Gd grade fell from 206,950 to 149,525. This lesson does the same column properly, and the change is one line.

Syntax: Series.fillna('Missing'). The label is arbitrary and any string that cannot collide with a real category works. The column stays text, so every encoder that follows treats the new label as one more level.

df['FireplaceQu'] = df['FireplaceQu'].fillna('Missing')   # one more level
df['GarageType'] = df['GarageType'].fillna('Missing')     # the same everywhere

Loading the house price data

These examples use the house price table and three of its columns: two text columns with missing values and the sale price to read the result against. Run this block first, then the example under it.

import pandas as pd
import matplotlib
matplotlib.use('svg')
import matplotlib.pyplot as plt

url = ("https://raw.githubusercontent.com/krishnaik06/"
       "Advanced-House-Price-Prediction-/master/train.csv")
df = pd.read_csv(url, usecols=['FireplaceQu', 'GarageType', 'SalePrice'])

Turning 690 gaps into a level

The fill is one call, and the result is a column with six levels instead of five. The new level is the largest one, which is honest: no fireplace is the most common thing to know about a house in this table.

df['FireplaceQu'] = df['FireplaceQu'].fillna('Missing')
print(df['FireplaceQu'].value_counts().to_string())
print()
print("levels now:", df['FireplaceQu'].nunique())
print("nulls left:", int(df['FireplaceQu'].isnull().sum()))
print("dtype:", df['FireplaceQu'].dtype)

FireplaceQu
Missing    690
Gd         380
TA         313
Fa          33
Ex          24
Po          20

levels now: 6
nulls left: 0
dtype: str

Reading the new level against the target

The test of whether the label deserves to exist is whether it behaves differently from the real categories. Grouping the sale price by the six levels answers it.

df['FireplaceQu'] = df['FireplaceQu'].fillna('Missing')
summary = (df.groupby('FireplaceQu')['SalePrice']
             .agg(['size', 'median']).sort_values('median'))
print(summary.round(0).to_string())

fig, ax = plt.subplots(figsize=(7, 4))
summary['median'].plot.bar(ax=ax, rot=0,
                           color=['#d64541' if i == 'Missing' else '#3a6fd8'
                                  for i in summary.index])
ax.set_ylabel('median SalePrice')
ax.set_title('Missing is a level of its own, and nearly the cheapest')
fig.savefig('chart.svg')

             size    median
FireplaceQu                
Po             20  131500.0
Missing       690  135000.0
Fa             33  158000.0
TA            313  187500.0
Gd            380  206950.0
Ex             24  314250.0

• Missing is 690 houses with a median price of 135,000. That is nearly 72,000 below the Gd grade and within 3,500 of Po, the cheapest real grade.
• The five real grades keep their prices. Nothing was poured into them, so Gd is still 206,950 and the ordering of the grades still makes sense.
• The absence turns out to mean something concrete. A missing fireplace grade in this table means the house has no fireplace, and houses without one are cheaper.
• The column is still text. It will be encoded like any other categorical column, with one extra level.

When a label beats a fill

Where this is the first thing to reach for

• A column whose gap has a meaning in the domain. No fireplace, no garage, no second phone number: absence is a fact, not a loss.
• A column with a large share missing. No fill value is honest over half a column, and a label costs nothing.
• Before any categorical encoder. One hot encoding and the target based encoders need a complete column, and this is the cheapest way to give them one, as One hot encoding assumes.

Related

• Previous: Frequent category imputation
• Next: SimpleImputer in scikit-learn
• See also: One hot encoding
• Reference: SimpleImputer with strategy='constant'