如果与前一行的差异高于阈值，则 Pandas 替换为 NaN

Question

I have an half hourly dataframe df from which i want to remove outliers.

date  = ['2015-02-03 23:00:00','2015-02-03 23:30:00','2015-02-04 00:00:00','2015-02-04 00:30:00']
value_column = [33.24  , 500  , 34.39  , 34.49 ]

df = pd.DataFrame({'value column':value_column})
df.index = pd.to_datetime(df['index'],format='%Y-%m-%d %H:%M')
df.drop(['index'],axis=1,inplace=True)

print(df.head())
                   value column  
index                                     
2015-02-03 23:00:00  33.24   
2015-02-03 23:30:00  500   
2015-02-04 00:00:00  34.39   
2015-02-04 00:30:00  34.49   

I want to remove outliers based on the difference of the values from one hour to the next. I would like to replace outliers values by NaN if the absolute difference from one hour to the next is above a given treshold. How can I do that efficiently?

I know that I can get the difference of the dataframe with the line below, however I do not know how to replace values by nan at the identified indexes where the difference is above the given treshold. Any idea on how to do that efficiently? (Assuming for instance that the treshold is 100)

df = df.diff()

I have tried the following, it does not throw any error but does not work:

df["value column"]=df["value column"].mask(df["value column"].diff().abs() > 100, np.nan) 

Expected results:

                   value column  
index                                     
2015-02-03 23:00:00  33.24   
2015-02-03 23:30:00  NaN   
2015-02-04 00:00:00  34.39   
2015-02-04 00:30:00  34.49   

Answer 1

You need to find the do diff from top and bot together or the row number 3 will be droped as well

df["value column"].mask((df["value column"].diff(-1).abs()>100) & (df["value column"].diff().abs() > 100), np.nan) 
Out[270]: 
0    33.24
1      NaN
2    34.39
3    34.49
Name: value column, dtype: float64

Answer 2

One strategy would be to append the df.diff() values as a new column to your dataframe and then use the df.apply() method in every row to return either the original row value or NaN depending on the value of the newly appended diff column. Keep in mind that the df.diff() will return NaN for the first row so you need to manually account for that in the "selection function" in your apply function.

df['diff'] = df.diff()
df['value column'] = df.apply(lambda x: x[0] if x[-1]<=100 or np.isnan(x[-1]) else np.nan , axis=1)
df

Results:

                     value column
index                            
2015-02-03 23:00:00         33.24
2015-02-03 23:30:00           NaN
2015-02-04 00:00:00         34.39
2015-02-04 00:30:00         34.49