如何优雅处理Pandas DataFrame指定列的非空值？

Great question! Your current loc approach gets the job done, but we can definitely make this more concise and idiomatic to Pandas. Let's look at a few better approaches:

1. Use where for Vectorized Value Replacement #

The where method is perfect here—it keeps values that meet a condition (in this case, null values) and replaces the rest with your desired expression. It's fully vectorized, so it's efficient and readable:

import pandas as pd

for col in columns_to_process:
    df[col] = df[col].where(df[col].isnull(), df[col].astype(str) + '!')

How it works: #

• df[col].isnull() identifies all positions with empty values, which where leaves untouched.
• For non-null positions, it runs df[col].astype(str) + '!' to append the exclamation mark.

2. Use mask (Reverse of where) #

If you prefer to start with the transformed values and then "restore" the nulls, mask is a clean alternative:

for col in columns_to_process:
    df[col] = (df[col].astype(str) + '!').mask(df[col].isnull())

How it works: #

• First, we convert all values to strings and add ! (including turning NaN to 'nan!' temporarily).
• Then mask replaces any positions where the original column had nulls back to NaN, preserving empty values as intended.

3. Avoid Loops Entirely with apply on Columns #

If you want to skip the explicit for loop, you can apply the logic to all target columns at once using df.apply:

df[columns_to_process] = df[columns_to_process].apply(
    lambda series: series.where(series.isnull(), series.astype(str) + '!')
)

This is even more Pythonic because it leverages Pandas' ability to operate on entire column groups instead of iterating one by one.

Why These Are Better Than Your Original loc Approach #

• Conciseness: All these methods cut down on repetitive code (no need to write df.loc[...] twice per column).
• Readability: The intent is immediately clear—either keep nulls and modify others, or modify first then restore nulls.
• Efficiency: Like loc, these are vectorized operations, so they're just as fast as your original solution (no performance tradeoff).

内容的提问来源于stack exchange，提问作者Rory LM