修改

Question

当我在pandas中有一个数据框时，如：

raw_data = {
        'subject_id': ['1', '2', '3', '4', '5'],
        'name': ['A', 'B', 'C', 'D', 'E'],
        'nationality': ['DE', 'AUT', 'US', 'US', 'US'],
        'alotdifferent': ['x', 'y', 'z', 'x', 'a'],
        'target': [0,0,0,1,1],
        'age_group' : [1, 2, 1, 3, 1]}
df_a = pd.DataFrame(raw_data, columns = ['subject_id', 'name', 'nationality', 'alotdifferent','target','age_group'])
df_a.nationality = df_a.nationality.astype('category')
df_a.alotdifferent = df_a.alotdifferent.astype('category')
df_a.name = df_a.name.astype('category')

目前，我使用：

FACTOR_FIELDS = df_a.select_dtypes(include=['category']).columns
columnsToDrop = ['alotdifferent']
columnsToBias_keep = FACTOR_FIELDS[~FACTOR_FIELDS.isin(columnsToDrop)]
target = 'target'

def quotients_slow(df_a):
    # parallelism = 8
    # original = dd.from_pandas(df.copy())
    original = df_a.copy()
    output_df = original
    ratio_weights = {}

    for colname in columnsToBias_keep.union(columnsToDrop):
        # group only a single time
        grouped = original.groupby([colname, target]).size() 
        # calculate first ratio
        df = grouped / original[target].sum() 
        nameCol = "pre_" + colname 
        grouped_res = df.reset_index(name=nameCol) 
        grouped_res = grouped_res[grouped_res[target] == 1] 
        grouped_res = grouped_res.drop(target, 1) 
        # todo persist the result in dict for transformer
        result_1 = grouped_res

        # calculate second ratio
        df = (grouped / grouped.groupby(level=0).sum()) 
        nameCol_2 = "pre2_" + colname 
        grouped = df.reset_index(name=nameCol_2) 
        grouped_res = grouped[grouped[target] == 1] 
        grouped_res = grouped_res.drop(target, 1) 
        result_2 = grouped_res 

        # persist the result in dict for transformer
        # this is required to separate fit and transform stage (later on in a sklearn transformer)
        ratio_weights[nameCol] = result_1
        ratio_weights[nameCol_2] = result_2

        # retrieve results
        res_1 = ratio_weights['pre_' + colname]
        res_2 = ratio_weights['pre2_' + colname]
        # merge ratio_weight with original dataframe
        output_df = pd.merge(output_df, res_1, on=colname, how='left') 
        output_df = pd.merge(output_df, res_2, on=colname, how='left') 
        output_df.loc[(output_df[nameCol].isnull()), nameCol] = 0 
        output_df.loc[(output_df[nameCol_2].isnull()), nameCol_2] = 0 

        if colname in columnsToDrop:
            output_df = output_df.drop(colname, 1)

    return output_df


quotients_slow(df_a)

以两种方式计算每个（分类）列的每个组与target:1的比率。由于我想对多列执行此操作，我天真地迭代所有这些操作。但是这个操作非常慢。在示例中：10 loops, best of 3: 37 ms per loop。对于我的大约500000行和大约100列的真实数据集，这确实需要一段时间。

是不是可以在dask或pandas中加速它（列并行方式，平凡的并行化）？是否有可能在普通大熊猫中更有效地实施它？是否可以减少用于计算商数的数据传递次数？

修改

尝试在for循环中使用dask.delayed来实现列的并行性时，我无法弄清楚如何在列上构建图形，因为我需要调用compute来获取元组。

delayed_res_name = delayed(compute_weights)(df_a, 'name')
a,b,c,d = delayed_res_name.compute()
ratio_weights = {}
ratio_weights[c] = a
ratio_weights[d] = b

Answer 1

这是使用Pandas为您的第一个商提供相当快速的解决方案。它假设您对计算subject_id的比例不感兴趣。我还在您的示例中添加了一些数据，以涵盖更多边缘情况。

首先，生成样本数据：

raw_data = {
    'subject_id': ['1', '2', '3', '4', '5', '6','7'],
    'name': ['A', 'B', 'C', 'D', 'E', 'A','A'],
    'nationality': ['DE', 'AUT', 'US', 'US', 'US', 'DE','DE'],
    'alotdifferent': ['x', 'y', 'z', 'x', 'a','x','z'],
    'target': [0,0,0,1,1,0,1],
    'age_group' : [1, 2, 1, 3, 1, 2,1]}

df_a = pd.DataFrame(raw_data, columns = ['subject_id', 'name', 'nationality', 'alotdifferent','target','age_group'])

现在计算比例并测量速度：

def compute_prop(group):
    return group.sum() / float(group.count())

def build_master(df):
    master = df.copy()
    fields = df.drop(['subject_id','target'],1).columns

    for field in fields:
        master = (pd.merge(master, df.groupby(field, as_index=False)
                                     .agg({'target':compute_prop})
                                     .rename(columns={'target':'pre_{}'.format(field)}), 
                           on=field)
             )

    master.sort_values('subject_id')
    return master

%timeit master = build_master(df_a)
10 loops, best of 3: 17.1 ms per loop

输出：

  subject_id name nationality alotdifferent  target  age_group  pre_name  \
0          1    A          DE             x       0          1  0.333333   
5          2    B         AUT             y       0          2  0.000000   
2          3    C          US             z       0          1  0.000000   
6          4    D          US             x       1          3  1.000000   
3          5    E          US             a       1          1  1.000000   
4          6    A          DE             x       0          2  0.333333   
1          7    A          DE             z       1          1  0.333333   

   pre_nationality  pre_alotdifferent  pre_age_group  
0         0.333333           0.333333            0.5  
5         0.000000           0.000000            0.0  
2         0.666667           0.500000            0.5  
6         0.666667           0.333333            1.0  
3         0.666667           1.000000            0.5  
4         0.333333           0.333333            0.0  
1         0.333333           0.500000            0.5

pandas / dask计算多列的百分比 - 列并行操作

修改

1 个答案: