更正 Pandas DataFrame 中的混乱日期

Question

我有一个百万行的时间序列数据框，其中 Date 列中的某些值具有混乱的日/月值。我如何有效地理清它们而又不破坏那些正确的东西？# this creates a dataframe with muddled datesimport pandas as pdimport numpy as npfrom pandas import Timestampstart = Timestamp(2013,1,1)dates = pd.date_range(start, periods=942)[::-1]muddler = {}for d in dates:    if d.day < 13:        muddler[d] = Timestamp(d.year, d.day, d.month)    else:        muddler[d] = Timestamp(d.year, d.month, d.day)df = pd.DataFrame()df['Date'] = datesdf['Date'] =  df['Date'].map(muddler)# now what? (assuming I don't know how the dates are muddled)

Answer 1

一个选项可能是计算时间戳的拟合度，并修改那些偏离拟合度大于特定阈值的时间戳。例子：

import pandas as pd

import numpy as np

start = pd.Timestamp(2013,1,1)

dates = pd.date_range(start, periods=942)[::-1]

muddler = {}

for d in dates:

    if d.day < 13:

        muddler[d] = pd.Timestamp(d.year, d.day, d.month)

    else:

        muddler[d] = pd.Timestamp(d.year, d.month, d.day)

df = pd.DataFrame()

df['Date'] = dates

df['Date'] =  df['Date'].map(muddler)

# convert date col to posix timestamp

df['ts'] = df['Date'].values.astype(np.float) / 10**9

# calculate a linear fit for ts col

x = np.linspace(df['ts'].iloc[0], df['ts'].iloc[-1], df['ts'].size)

df['ts_linfit'] = np.polyval(np.polyfit(x, df['ts'], 1), x)

# set a thresh and derive a mask that masks differences between 

# fit and timestamp greater than thresh:

thresh = 1.2e6 # you might want to tweak this...

m = np.absolute(df['ts']-df['ts_linfit']) > thresh

# create new date col as copy of original

df['Date_filtered'] = df['Date']

# modify values that were caught in the mask

df.loc[m, 'Date_filtered'] = df['Date_filtered'][m].apply(lambda x: pd.Timestamp(x.year, x.day, x.month))

# also to posix timestamp

df['ts_filtered'] = df['Date_filtered'].values.astype(np.float) / 10**9

ax = df['ts'].plot(label='original')

ax = df['ts_filtered'].plot(label='filtered')

ax.legend()

Answer 2

在尝试创建一个最小的可重现示例时，我实际上已经解决了我的问题——但我希望有一种更有效的方法来做我想做的事情……

# i first define a function to examine the dates

def disordered_muddle(date_series, future_first=True):

    """Check whether a series of dates is disordered or just muddled"""

    disordered = []

    muddle = []

    dates = date_series

    different_dates = pd.Series(dates.unique())

    date = different_dates[0]

    for i, d in enumerate(different_dates[1:]):

        # we expect the date's dayofyear to decrease by one

        if d.dayofyear!=date.dayofyear-1:

            # unless the year is changing

            if d.year!=date.year-1:

                try:

                    # we check if the day and month are muddled

                    # if d.day > 12 this will cause an Exception

                    unmuddle = Timestamp(d.year,d.day,d.month)

                    if unmuddle.dayofyear==date.dayofyear-1:

                        muddle.append(d)

                        d = unmuddle

                    elif unmuddle.year==date.year-1:

                        muddle.append(d)

                        d = unmuddle

                    else:

                        disordered.append(d)

                except:

                    disordered.append(d)

        date=d

    if len(disordered)==0 and len(muddle)==0:

        return False

    else:

        return disordered, muddle

disorder, muddle = disordered_muddle(df['Date'])

# finally unmuddle the dates

date_correction = {}

for d in df['Date']:

    if d in muddle:

        date_correction[d] = Timestamp(d.year, d.day, d.month)

    else:

        date_correction[d] = Timestamp(d.year, d.month, d.day)

df['CorrectedDate'] = df['Date'].map(date_correction)

disordered_muddle(df['CorrectedDate'])