Handling Missing Data

title: Handling Missing Data keywords: [handling missing data, handling missing data in python, python data analysis, dataidea, machine learning, data preprocessing, sklearn data preprocessing] description: n this notebook, we look into the top four missing data imputation methods author: Juma Shafara date: "2024-03"

Photo by DATAIDEA

Introduction:

Missing data is a common hurdle in data analysis, impacting the reliability of insights drawn from datasets. Python offers a range of solutions to address this issue, some of which we discussed in the earlier weeks. In this notebook, we look into the top four missing data imputation methods:

SimpleImputer
KNNImputer
IterativeImputer
Datawig

We'll explore these essential techniques, using sklearn and the weather dataset.

# install the libraries for this demonstration
# ! pip install -U dataidea

import pandas as pd
import dataidea as di

loadDataset allows us to load datasets inbuilt in the dataidea library

weather = di.loadDataset('weather') 
weather

	day	temperature	windspead	event
0	01/01/2017	32.0	6.0	Rain
1	04/01/2017	NaN	9.0	Sunny
2	05/01/2017	28.0	NaN	Snow
3	06/01/2017	NaN	7.0	NaN
4	07/01/2017	32.0	NaN	Rain
5	08/01/2017	NaN	NaN	Sunny
6	09/01/2017	NaN	NaN	NaN
7	10/01/2017	34.0	8.0	Cloudy
8	11/01/2017	40.0	12.0	Sunny

weather.isna().sum()

day            0
temperature    4
windspead      4
event          2
dtype: int64

Let's demonstrate how to use the top three missing data imputation methods—SimpleImputer, KNNImputer, and IterativeImputer—using the simple weather dataset.

# select age from the data
temp_wind = weather[['temperature', 'windspead']].copy()

temp_wind_imputed = temp_wind.copy()

SimpleImputer from scikit-learn:

Usage: SimpleImputer is a straightforward method for imputing missing values by replacing them with a constant, mean, median, or most frequent value along each column.
Pros:
- Easy to use and understand.
- Can handle both numerical and categorical data.
- Offers flexibility with different imputation strategies.
Cons:
- It doesn't consider relationships between features.
- May not be the best choice for datasets with complex patterns of missingness.
Example:

from sklearn.impute import SimpleImputer

simple_imputer = SimpleImputer(strategy='mean')
temp_wind_simple_imputed = simple_imputer.fit_transform(temp_wind)

temp_wind_simple_imputed_df = pd.DataFrame(temp_wind_simple_imputed, columns=temp_wind.columns)

Let's have a look at the outcome

temp_wind_simple_imputed_df

	temperature	windspead
0	32.0	6.0
1	33.2	9.0
2	28.0	8.4
3	33.2	7.0
4	32.0	8.4
5	33.2	8.4
6	33.2	8.4
7	34.0	8.0
8	40.0	12.0

Exercise:

Try out the SimpleImputer with different imputation strategies like mode, constant
Choose and try some imputation techniques on categorical data

KNNImputer from scikit-learn:

Usage:
- KNNImputer imputes missing values using k-nearest neighbors, replacing them with the mean value of the nearest neighbors.
- You can read more about the KNNImputer from the sklearn official docs site
Pros:
- Considers relationships between features, making it suitable for datasets with complex patterns of missingness.
- Can handle both numerical and categorical data.
Cons:
- Computationally expensive for large datasets.
- Requires careful selection of the number of neighbors (k).

Note!

By default, the KNNImputer uses 'nan' values as missing data and the 'nan_euclidean' metric to calculate the distances between values.

Example:

from sklearn.impute import KNNImputer

knn_imputer = KNNImputer(n_neighbors=2)
temp_wind_knn_imputed = knn_imputer.fit_transform(temp_wind)

temp_wind_knn_imputed_df = pd.DataFrame(temp_wind_knn_imputed, columns=temp_wind.columns)

If we take a look at the outcome

weather

	day	temperature	windspead	event
0	01/01/2017	32.0	6.0	Rain
1	04/01/2017	NaN	9.0	Sunny
2	05/01/2017	28.0	NaN	Snow
3	06/01/2017	NaN	7.0	NaN
4	07/01/2017	32.0	NaN	Rain
5	08/01/2017	NaN	NaN	Sunny
6	09/01/2017	NaN	NaN	NaN
7	10/01/2017	34.0	8.0	Cloudy
8	11/01/2017	40.0	12.0	Sunny

Filling a single column independently using the `KNNImputer`

To use the KNNImputer for a single independ column, you can use the index as the other column instead, this will result into equal euclidean distances resulting into the use of the physical neighbors in the data table.

from sklearn.impute import KNNImputer

knn_imputer = KNNImputer(n_neighbors=2)
windspead_imputed = knn_imputer.fit_transform(weather[['windspead']].reset_index())

windspead_imputed

array([[ 0. ,  6. ],
       [ 1. ,  9. ],
       [ 2. ,  8. ],
       [ 3. ,  7. ],
       [ 4. ,  8. ],
       [ 5. ,  7.5],
       [ 6. , 10. ],
       [ 7. ,  8. ],
       [ 8. , 12. ]])

# we can fill it back in the weather data
weather['windspead'] = windspead_imputed[:, 1]

# now looking at the data
weather

	day	temperature	windspead	event
0	01/01/2017	32.0	6.0	Rain
1	04/01/2017	NaN	9.0	Sunny
2	05/01/2017	28.0	8.0	Snow
3	06/01/2017	NaN	7.0	NaN
4	07/01/2017	32.0	8.0	Rain
5	08/01/2017	NaN	7.5	Sunny
6	09/01/2017	NaN	10.0	NaN
7	10/01/2017	34.0	8.0	Cloudy
8	11/01/2017	40.0	12.0	Sunny

Exercise

Try out the KNNImputer with different numbers of neighbors and compare the results
Findo out how to use KNNImputer to fill categorical data

IterativeImputer from scikit-learn:

Usage: IterativeImputer models each feature with missing values as a function of other features and uses that estimate for imputation. It iteratively estimates the missing values.
Pros:
- Takes into account relationships between features, making it suitable for datasets with complex missing patterns.
- More robust than SimpleImputer for handling missing data.
Cons:
- Can be computationally intensive and slower than SimpleImputer.
- Requires careful tuning of model parameters.
Example:

from sklearn.experimental import enable_iterative_imputer
from sklearn.impute import IterativeImputer

iterative_imputer = IterativeImputer()
temp_wind_iterative_imputed = iterative_imputer.fit_transform(temp_wind)

temp_wind_iterative_imputed_df = pd.DataFrame(temp_wind_iterative_imputed, columns=temp_wind.columns)

temp_wind_iterative_imputed_df

	temperature	windspead
0	32.000000	6.0
1	33.967053	9.0
2	28.000000	8.0
3	31.410210	7.0
4	32.000000	8.0
5	32.049421	7.5
6	35.245474	10.0
7	34.000000	8.0
8	40.000000	12.0

You can also choose an estimator of your choice, let's try a Linear Regression model

from sklearn.linear_model import LinearRegression
from sklearn.experimental import enable_iterative_imputer
from sklearn.impute import IterativeImputer

# set estimator to an instance of a model
iterative_imputer = IterativeImputer(estimator=LinearRegression())
temp_wind_iterative_imputed = iterative_imputer.fit_transform(temp_wind)

temp_wind_iterative_imputed_df = pd.DataFrame(temp_wind_iterative_imputed, columns=temp_wind.columns)

temp_wind_iterative_imputed_df

	temperature	windspead
0	32.000000	6.0
1	34.125000	9.0
2	28.000000	8.0
3	31.041667	7.0
4	32.000000	8.0
5	31.812500	7.5
6	35.666667	10.0
7	34.000000	8.0
8	40.000000	12.0

Datawig:

Datawig is a library specifically designed for imputing missing values in tabular data using deep learning models.

# import datawig

# # Impute missing values
# df_imputed = datawig.SimpleImputer.complete(weather)

These top imputation methods offer different trade-offs in terms of computational complexity, handling of missing data patterns, and ease of use. The choice between them depends on the specific characteristics of the dataset and the requirements of the analysis.

Homework

Try out these techniques for categorical data

Don't miss out on any updates and developments! Subscribe to the DATAIDEA Newsletter it's easy and safe.