计算pandas中groupby的nunique()

Question

我有一个带有列的数据框：

1. diff - 注册日期和付款日期之间的差异，以天为单位
2. country - 用户所在国家/地区
3. user_id
4. campaign_id -- 另一个分类列，我们将在 groupby 中使用它

我需要为每个拥有diffcountry+campaign_id 组计算不同用户的数量。 例如，对于country'A'、campaign'abc'和diff7，我需要从country'A'、campaign'abc'和diff中获取不同的用户数 7

我当前的解决方案（如下）工作时间过长

import pandas as pd
import numpy as np

## generate test dataframe
df = pd.DataFrame({
        'country':np.random.choice(['A', 'B', 'C', 'D'], 10000),
        'campaign': np.random.choice(['camp1', 'camp2', 'camp3', 'camp4', 'camp5', 'camp6'], 10000),
        'diff':np.random.choice(range(10), 10000),
        'user_id': np.random.choice(range(1000), 10000)
        })
## main
result_df = pd.DataFrame()
for diff in df['diff'].unique():
    tmp_df = df.loc[df['diff']<=diff,:]
    tmp_df = tmp_df.groupby(['country', 'campaign'], as_index=False).apply(lambda x: x.user_id.nunique()).reset_index()
    tmp_df['diff'] = diff
    tmp_df.columns=['country', 'campaign', 'unique_ppl', 'diff']
    result_df = pd.concat([result_df, tmp_df],ignore_index=True, axis=0)

也许有更好的方法来做到这一点？

Answer 1

首先使用concat 和assign 的列表理解将所有连接在一起，然后使用groupby 和nunique 添加列diff，最后重命名列，如果需要，添加reindex 用于自定义列顺序：

df1 = pd.concat([df.loc[df['diff']<=x].assign(diff=x) for x in  df['diff'].unique()])
df2 = (df1.groupby(['diff','country', 'campaign'], sort=False)['user_id']
          .nunique()
          .reset_index()
          .rename(columns={'user_id':'unique_ppl'})
          .reindex(columns=['country', 'campaign', 'unique_ppl', 'diff']))

Answer 2

下面有一个替代方案，但@jezrael's solution 是最佳选择。

性能基准测试

%timeit original(df)  # 149ms
%timeit jp(df)        # 81ms
%timeit jez(df)       # 47ms

def original(df):
    result_df = pd.DataFrame()
    for diff in df['diff'].unique():
        tmp_df = df.loc[df['diff']<=diff,:]
        tmp_df = tmp_df.groupby(['country', 'campaign'], as_index=False).apply(lambda x: x.user_id.nunique()).reset_index()
        tmp_df['diff'] = diff
        tmp_df.columns=['country', 'campaign', 'unique_ppl', 'diff']
        result_df = pd.concat([result_df, tmp_df],ignore_index=True, axis=0)

    return result_df

def jp(df):

    result_df = pd.DataFrame()
    lst = []
    lst_append = lst.append
    for diff in df['diff'].unique():
        tmp_df = df.loc[df['diff']<=diff,:]
        tmp_df = tmp_df.groupby(['country', 'campaign'], as_index=False).agg({'user_id': 'nunique'})
        tmp_df['diff'] = diff
        tmp_df.columns=['country', 'campaign', 'unique_ppl', 'diff']
        lst_append(tmp_df)

    result_df = result_df.append(pd.concat(lst, ignore_index=True, axis=0), ignore_index=True)

    return result_df

def jez(df):
    df1 = pd.concat([df.loc[df['diff']<=x].assign(diff=x) for x in  df['diff'].unique()])
    df2 = (df1.groupby(['diff','country', 'campaign'], sort=False)['user_id']
              .nunique()
              .reset_index()
              .rename(columns={'user_id':'unique_ppl'})
              .reindex(columns=['country', 'campaign', 'unique_ppl', 'diff']))
    return df2