Python嵌套循环仅保留最后轮结果：DataFrame聚类集标记异常

It looks like your issue stems from not creating an explicit copy of your DataFrame slice and potentially running into pandas' SettingWithCopy behavior, where changes to a view of the original DataFrame don't persist correctly. Additionally, initializing your new columns upfront can help avoid unexpected behavior when assigning values.

Here's the fixed approach:

Step 1: Create a copy of your DataFrame and initialize new columns #

First, make sure you're working with a copy of the relevant columns (not just a view), and set up the cluster and cluster set columns with default values:

import pandas as pd

# Create a copy of the relevant columns to avoid SettingWithCopy issues
df_triplets = df[['Activity', 'SMILES']].copy()

# Initialize new columns with nullable values
df_triplets['cluster'] = pd.NA
df_triplets['cluster set'] = pd.NA

Step 2: Refactor the loop for better performance and correctness #

Instead of nesting three loops (which is slow for large datasets), you can use pandas' vectorized indexing to assign values to all molecules in a cluster at once. This is more efficient and avoids potential issues with individual row assignments:

list_points = [train_points, test_points, val_points]
name_points = ['train', 'test', 'val']

for name, points in zip(name_points, list_points):
    for cluster_num, cluster_indices in enumerate(points):
        # Assign cluster number to all molecules in the cluster
        df_triplets.loc[cluster_indices, 'cluster'] = cluster_num
        # Assign cluster set to all molecules in the cluster
        df_triplets.loc[cluster_indices, 'cluster set'] = name

Why this works: #

• Using .copy() ensures you're modifying a separate DataFrame, not a view of the original one, so changes persist correctly across all loop iterations.
• Initializing columns upfront means pandas doesn't have to dynamically add them during the loop, which can cause inconsistent behavior.
• Vectorized assignments (using a list of indices in loc) are significantly faster than looping through each molecule individually, especially with large datasets.

Verify the results #

You can confirm all sets are properly assigned with:

print(df_triplets['cluster set'].value_counts())

This should show counts for train, test, and val instead of just val.

内容的提问来源于stack exchange，提问作者Daniel