如何从DataFrame列的各分组中选取Top 100及Top N值？

Hey there! Let's break down these two common Pandas grouping tasks clearly, with practical code examples you can copy-paste and adjust for your use case.

If you just want to keep the first 100 records in their original order for each group, Pandas' groupby().head() method is the most hassle-free solution—it automatically takes all records if a group has fewer than 100 entries, no errors thrown.

Let's use a sample DataFrame with user behavior data as an example:

import pandas as pd
import numpy as np

# Build sample data
df = pd.DataFrame({
    'user_id': np.repeat(range(10), 150),  # 10 users, 150 records each
    'action_time': pd.date_range(start='2024-01-01', periods=1500, freq='min'),
    'action_type': np.random.choice(['click', 'scroll', 'purchase'], 1500)
})

# Group by user_id and take first 100 records per group
top_100_per_group = df.groupby('user_id').head(100)

If you need to sort by a column first (like reverse chronological order) before taking the top 100, just add a sort step:

# Sort by user_id, then reverse action_time, then take top 100 per group
sorted_top_100 = df.sort_values(['user_id', 'action_time'], ascending=[True, False])\
                   .groupby('user_id')\
                   .head(100)

When you need to pick the Top N values (largest or smallest) based on a numeric column for each group, here are two go-to methods:

Method 1: Use nlargest()/nsmallest() (intuitive and efficient) #

nlargest(N, column) directly returns the N records with the highest values in the specified column, while nsmallest() does the opposite. Pair it with groupby.apply() for clean code:

# Add a numeric column (e.g., click counts) to the sample data
df['click_count'] = np.random.randint(1, 100, 1500)

# Group by user_id, take Top 5 records with highest click_count per group
top_5_click = df.groupby('user_id').apply(lambda x: x.nlargest(5, 'click_count')).reset_index(drop=True)

The reset_index(drop=True) removes the multi-level index generated by grouping, making the result easier to work with.

Method 2: Sort then take top N (great for multi-column sorting) #

This method is more flexible if you need to sort by multiple columns before picking Top N:

# Sort by user_id, then descending click_count, then ascending action_time; take Top 5 per group
multi_sort_top_5 = df.sort_values(['user_id', 'click_count', 'action_time'], 
                                  ascending=[True, False, True])\
                     .groupby('user_id')\
                     .head(5)

Each method has its edge: nlargest() is concise for single-metric Top picks, while sorting then taking head handles complex sorting rules (like prioritizing value first, then time) smoothly.

内容的提问来源于stack exchange，提问作者Karishma Rastogi