Navigation

How to Handle Missing Data with Python

By on March 20, 2017 in Data Preparation

Last Updated on August 28, 2020

Real-world data often has missing values.

Data can have missing values for a number of reasons such as observations that were not recorded and data corruption.

Handling missing data is important as many machine learning algorithms do not support data with missing values.

In this tutorial, you will discover how to handle missing data for machine learning with Python.

Specifically, after completing this tutorial you will know:

  • How to marking invalid or corrupt values as missing in your dataset.
  • How to remove rows with missing data from your dataset.
  • How to impute missing values with mean values in your dataset.

Kick-start your project with my new book Data Preparation for Machine Learning, including step-by-step tutorials and the Python source code files for all examples.

Let’s get started.

Note: The examples in this post assume that you have Python 3 with Pandas, NumPy and Scikit-Learn installed, specifically scikit-learn version 0.22 or higher. If you need help setting up your environment see this tutorial.

  • Update Mar/2018: Changed link to dataset files.
  • Update Dec/2019: Updated link to dataset to GitHub version.
  • Update May/2020: Updated code examples for API changes. Added references.
How to Handle Missing Values with Python

How to Handle Missing Values with Python
Photo by CoCreatr, some rights reserved.

Overview

This tutorial is divided into 6 parts:

  1. Diabetes Dataset: where we look at a dataset that has known missing values.
  2. Mark Missing Values: where we learn how to mark missing values in a dataset.
  3. Missing Values Causes Problems: where we see how a machine learning algorithm can fail when it contains missing values.
  4. Remove Rows With Missing Values: where we see how to remove rows that contain missing values.
  5. Impute Missing Values: where we replace missing values with sensible values.
  6. Algorithms that Support Missing Values: where we learn about algorithms that support missing values.

First, let’s take a look at our sample dataset with missing values.

1. Diabetes Dataset

The Diabetes Dataset involves predicting the onset of diabetes within 5 years in given medical details.

It is a binary (2-class) classification problem. The number of observations for each class is not balanced. There are 768 observations with 8 input variables and 1 output variable. The variable names are as follows:

  • 0. Number of times pregnant.
  • 1. Plasma glucose concentration a 2 hours in an oral glucose tolerance test.
  • 2. Diastolic blood pressure (mm Hg).
  • 3. Triceps skinfold thickness (mm).
  • 4. 2-Hour serum insulin (mu U/ml).
  • 5. Body mass index (weight in kg/(height in m)^2).
  • 6. Diabetes pedigree function.
  • 7. Age (years).
  • 8. Class variable (0 or 1).

The baseline performance of predicting the most prevalent class is a classification accuracy of approximately 65%. Top results achieve a classification accuracy of approximately 77%.

A sample of the first 5 rows is listed below.

This dataset is known to have missing values.

Specifically, there are missing observations for some columns that are marked as a zero value.

We can corroborate this by the definition of those columns and the domain knowledge that a zero value is invalid for those measures, e.g. a zero for body mass index or blood pressure is invalid.

Download the dataset from here and save it to your current working directory with the file name pima-indians-diabetes.csv .

Want to Get Started With Data Preparation?

Take my free 7-day email crash course now (with sample code).

Click to sign-up and also get a free PDF Ebook version of the course.

Download Your FREE Mini-Course

2. Mark Missing Values

Most data has missing values, and the likelihood of having missing values increases with the size of the dataset.

Missing data are not rare in real data sets. In fact, the chance that at least one data point is missing increases as the data set size increases.

— Page 187, Feature Engineering and Selection, 2019.

In this section, we will look at how we can identify and mark values as missing.

We can use plots and summary statistics to help identify missing or corrupt data.

We can load the dataset as a Pandas DataFrame and print summary statistics on each attribute.

Running this example produces the following output:

This is useful.

We can see that there are columns that have a minimum value of zero (0). On some columns, a value of zero does not make sense and indicates an invalid or missing value.

Missing values are frequently indicated by out-of-range entries; perhaps a negative number (e.g., -1) in a numeric field that is normally only positive, or a 0 in a numeric field that can never normally be 0.

— Page 62, Data Mining: Practical Machine Learning Tools and Techniques, 2016.

Specifically, the following columns have an invalid zero minimum value:

  • 1: Plasma glucose concentration
  • 2: Diastolic blood pressure
  • 3: Triceps skinfold thickness
  • 4: 2-Hour serum insulin
  • 5: Body mass index

Let’s confirm this my looking at the raw data, the example prints the first 20 rows of data.

Running the example, we can clearly see 0 values in the columns 2, 3, 4, and 5.

We can get a count of the number of missing values on each of these columns. We can do this my marking all of the values in the subset of the DataFrame we are interested in that have zero values as True. We can then count the number of true values in each column.

We can do this my marking all of the values in the subset of the DataFrame we are interested in that have zero values as True. We can then count the number of true values in each column.

Running the example prints the following output:

We can see that columns 1,2 and 5 have just a few zero values, whereas columns 3 and 4 show a lot more, nearly half of the rows.

This highlights that different “missing value” strategies may be needed for different columns, e.g. to ensure that there are still a sufficient number of records left to train a predictive model.

When a predictor is discrete in nature, missingness can be directly encoded into the predictor as if it were a naturally occurring category.

— Page 197, Feature Engineering and Selection, 2019.

In Python, specifically Pandas, NumPy and Scikit-Learn, we mark missing values as NaN.

Values with a NaN value are ignored from operations like sum, count, etc.

We can mark values as NaN easily with the Pandas DataFrame by using the replace() function on a subset of the columns we are interested in.

After we have marked the missing values, we can use the isnull() function to mark all of the NaN values in the dataset as True and get a count of the missing values for each column.

Running the example prints the number of missing values in each column. We can see that the columns 1:5 have the same number of missing values as zero values identified above. This is a sign that we have marked the identified missing values correctly.

We can see that the columns 1 to 5 have the same number of missing values as zero values identified above. This is a sign that we have marked the identified missing values correctly.

This is a useful summary. I always like to look at the actual data though, to confirm that I have not fooled myself.

Below is the same example, except we print the first 20 rows of data.

Running the example, we can clearly see NaN values in the columns 2, 3, 4 and 5. There are only 5 missing values in column 1, so it is not surprising we did not see an example in the first 20 rows.

It is clear from the raw data that marking the missing values had the intended effect.

Before we look at handling missing values, let’s first demonstrate that having missing values in a dataset can cause problems.

3. Missing Values Causes Problems

Having missing values in a dataset can cause errors with some machine learning algorithms.

Missing values are common occurrences in data. Unfortunately, most predictive modeling techniques cannot handle any missing values. Therefore, this problem must be addressed prior to modeling.

— Page 203, Feature Engineering and Selection, 2019.

In this section, we will try to evaluate a the Linear Discriminant Analysis (LDA) algorithm on the dataset with missing values.

This is an algorithm that does not work when there are missing values in the dataset.

The below example marks the missing values in the dataset, as we did in the previous section, then attempts to evaluate LDA using 3-fold cross validation and print the mean accuracy.

Running the example results in an error, as follows:

This is as we expect.

We are prevented from evaluating an LDA algorithm (and other algorithms) on the dataset with missing values.

Many popular predictive models such as support vector machines, the glmnet, and neural networks, cannot tolerate any amount of missing values.

— Page 195, Feature Engineering and Selection, 2019.

Now, we can look at methods to handle the missing values.

4. Remove Rows With Missing Values

The simplest strategy for handling missing data is to remove records that contain a missing value.

The simplest approach for dealing with missing values is to remove entire predictor(s) and/or sample(s) that contain missing values.

— Page 196, Feature Engineering and Selection, 2019.

We can do this by creating a new Pandas DataFrame with the rows containing missing values removed.

Pandas provides the dropna() function that can be used to drop either columns or rows with missing data. We can use dropna() to remove all rows with missing data, as follows:

Running this example, we can see that the number of rows has been aggressively cut from 768 in the original dataset to 392 with all rows containing a NaN removed.

We now have a dataset that we could use to evaluate an algorithm sensitive to missing values like LDA.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

The example runs successfully and prints the accuracy of the model.

Removing rows with missing values can be too limiting on some predictive modeling problems, an alternative is to impute missing values.

5. Impute Missing Values

Imputing refers to using a model to replace missing values.

… missing data can be imputed. In this case, we can use information in the training set predictors to, in essence, estimate the values of other predictors.

— Page 42, Applied Predictive Modeling, 2013.

There are many options we could consider when replacing a missing value, for example:

  • A constant value that has meaning within the domain, such as 0, distinct from all other values.
  • A value from another randomly selected record.
  • A mean, median or mode value for the column.
  • A value estimated by another predictive model.

Any imputing performed on the training dataset will have to be performed on new data in the future when predictions are needed from the finalized model. This needs to be taken into consideration when choosing how to impute the missing values.

For example, if you choose to impute with mean column values, these mean column values will need to be stored to file for later use on new data that has missing values.

Pandas provides the fillna() function for replacing missing values with a specific value.

For example, we can use fillna() to replace missing values with the mean value for each column, as follows:

Running the example provides a count of the number of missing values in each column, showing zero missing values.

The scikit-learn library provides the SimpleImputer pre-processing class that can be used to replace missing values.

It is a flexible class that allows you to specify the value to replace (it can be something other than NaN) and the technique used to replace it (such as mean, median, or mode). The SimpleImputer class operates directly on the NumPy array instead of the DataFrame.

The example below uses the SimpleImputer class to replace missing values with the mean of each column then prints the number of NaN values in the transformed matrix.

Running the example shows that all NaN values were imputed successfully.

In either case, we can train algorithms sensitive to NaN values in the transformed dataset, such as LDA.

The example below shows the LDA algorithm trained in the SimpleImputer transformed dataset.

We use a Pipeline to define the modeling pipeline, where data is first passed through the imputer transform, then provided to the model. This ensures that the imputer and model are both fit only on the training dataset and evaluated on the test dataset within each cross-validation fold. This is important to avoid data leakage.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

Running the example prints the accuracy of LDA on the transformed dataset.

Try replacing the missing values with other values and see if you can lift the performance of the model.

Maybe missing values have meaning in the data.

For a more detailed example of imputing missing values with statistics see the tutorial:

Next we will look at using algorithms that treat missing values as just another value when modeling.

6. Algorithms that Support Missing Values

Not all algorithms fail when there is missing data.

There are algorithms that can be made robust to missing data, such as k-Nearest Neighbors that can ignore a column from a distance measure when a value is missing. Naive Bayes can also support missing values when making a prediction.

One of the really nice things about Naive Bayes is that missing values are no problem at all.

— Page 100, Data Mining: Practical Machine Learning Tools and Techniques, 2016.

There are also algorithms that can use the missing value as a unique and different value when building the predictive model, such as classification and regression trees.

… a few predictive models, especially tree-based techniques, can specifically account for missing data.

— Page 42, Applied Predictive Modeling, 2013.

Sadly, the scikit-learn implementations of naive bayes, decision trees and k-Nearest Neighbors are not robust to missing values. Although it is being considered.

Nevertheless, this remains as an option if you consider using another algorithm implementation (such as xgboost) or developing your own implementation.

Further Reading

This section provides more resources on the topic if you are looking to go deeper.

Related Tutorials

Books

APIs

Summary

In this tutorial, you discovered how to handle machine learning data that contains missing values.

Specifically, you learned:

  • How to mark missing values in a dataset as numpy.nan.
  • How to remove rows from the dataset that contain missing values.
  • How to replace missing values with sensible values.

Do you have any questions about handling missing values?
Ask your questions in the comments and I will do my best to answer.

Get a Handle on Modern Data Preparation!

Data Preparation for Machine Learning

Prepare Your Machine Learning Data in Minutes

...with just a few lines of python code

Discover how in my new Ebook:
Data Preparation for Machine Learning

It provides self-study tutorials with full working code on:
Feature Selection, RFE, Data Cleaning, Data Transforms, Scaling, Dimensionality Reduction, and much more...

Bring Modern Data Preparation Techniques to
Your Machine Learning Projects


See What's Inside

126 Responses to How to Handle Missing Data with Python

  1. Mike March 20, 2017 at 3:16 pm #

    Fancy impute is a library i’ve turned too for imputation:

    https://github.com/hammerlab/fancyimpute

    Also missingno is great for visualizations!

    https://github.com/ResidentMario/missingno

    Reply
    • Jason Brownlee March 21, 2017 at 8:37 am #

      Thanks for the tip Mike.

      Reply
      • ishtiaq ahmed December 10, 2019 at 4:28 am #

        Hi, friend I need that dataset ” Pima-Indians-diabetes.csv” how can I access it. it is not available on this site

        Reply
        • Jason Brownlee December 10, 2019 at 7:35 am #

          All datasets are here:
          https://github.com/jbrownlee/Datasets

          Reply
          • ishtiaq ahmed December 11, 2019 at 2:00 am #

            thnx Jason

          • Jason Brownlee December 11, 2019 at 7:02 am #

            You’re welcome.

          • umer January 10, 2020 at 8:06 pm #

            Hi,
            I have a data set with 3 lakhs row and 278 columns. I used MissForest to impute missing values. But, the system (HP Pavilion Intel i5 with 12GB RAM) runs for a long time and still didn’t complete..Can you suggest any easy way?
            should I have to use any loop?

          • Jason Brownlee January 11, 2020 at 7:24 am #

            Perhaps use less data?
            Perhaps fit on a faster machine?

    • bakyalakshmi September 27, 2017 at 2:56 pm #

      please tell me about how to impute median using one dataset

      Reply
    • Trung Nguyen Thanh July 8, 2018 at 8:42 pm #

      please tell me, in case use Fancy impute library, how to predict for X_test?

      Reply
  2. Jozo Kovac April 1, 2017 at 8:06 am #

    Thanks for pointing on interesting problem. I would love another one about how to deal with categorical attributes in Python.

    And dear reader, please never ever remove rows with missing values. It changes the distribution of your data and your analyses may become worthless. Learn from mistakes of others and don’t repeat them 🙂

    Reply
  3. Tommy Carstensen April 4, 2017 at 3:56 am #

    Super duper! Thanks for writing! Would it have been worth mentioning interpolate of Pandas? http://pandas.pydata.org/pandas-docs/stable/generated/pandas.Series.interpolate.html

    Reply
  4. Aswathy April 14, 2017 at 12:10 pm #

    Hi Jason,

    I was just wondering if there is a way to use a different imputation strategy for each column. Say, for a categorical feature you want to impute using the mode but for a continuous attribute, you want to impute using mean.

    Reply
    • Jason Brownlee April 15, 2017 at 9:30 am #

      Yes, try lots of techniques, go with whatever results in the most accurate models.

      Reply
  5. Salu Khadka June 11, 2017 at 12:29 am #

    thanks for your tutorial sir.
    I would also seek help from you for multi label classification of a textual data , if possible.

    For example, categorizing a twitter post as related to sports, business , tech , or others.

    Reply
  6. Ali Gabriel Lara June 13, 2017 at 4:51 am #

    Hello Mr. Brownlee. Thank you so much for your post.

    Do you know any approach to recognize the pattern of missing data? I mean, I am interested in discovering the pattern of missing data on a time series data.

    The database is historical data of a chemical process. I think I should apply some pattern recognition approach columnwise because each column represents a process variable and the value coming from a transmisor.

    My goal is to predict if the missing data is for a mechanical fault or a desviation in registration process or for any other causes. Then I should apply a kind of filling methods if it is required.

    Have you any advice? Thanks in advance

    Reply
    • Jason Brownlee June 13, 2017 at 8:25 am #

      I would invert the problem and model the series of missing data and mark all data you do have with a special value “0” and all missing instances as “1”.

      Great problem!

      Let me know how you go.

      Reply
  7. Patricia Villa October 5, 2017 at 3:45 pm #

    You helped me keep my sanity. THANK YOU!!

    Reply
  8. Sachin Raj October 6, 2017 at 7:58 pm #

    How to know whether to apply mean or to replace it with mode?

    Reply
    • Jason Brownlee October 7, 2017 at 5:54 am #

      Try both and see what results in the most skillful models.

      Reply
    • Naga May 24, 2018 at 10:35 pm #

      Hi Sachin,
      Mode is effected by outliers whereas Mean is less effected by outliers.
      Please correct me if i am wrong@Jason

      Reply
      • Jason Brownlee May 25, 2018 at 9:26 am #

        I think you meant “Median” is not affected by outliers. “Mode” is just the most common value.

        Reply
  9. Patrick October 26, 2017 at 5:06 am #

    If I have a 11×11 table and there are 20 missing values in there, is there a way for me to make a code that creates a list after identifying these values?

    Let us say that the first column got names and the first row has Day 1 to 10. Some of the names does not show up all of the days and therefore there are missing gaps. I put this table into the code and rather than reading the table I get a list with:

    Name, day 2, day 5, day 7
    Name, Day 1, day 6

    I understand that this could take some time to answer, but if you are able to just tell me that this is possible and maybe know of good place to start on how to start on this project that would be of great help!

    Reply
    • Jason Brownlee October 26, 2017 at 5:35 am #

      Sure, if the missing values are marked with a nan or similar, you can retrieve rows with missing values using Pandas.

      Reply
  10. Nivetha December 22, 2017 at 7:46 pm #

    can we code our own algorithms to impute the missing values????
    if it is possible then how can i implement it??

    Reply
    • Jason Brownlee December 23, 2017 at 5:15 am #

      Yes.

      You can write some if-statements and fill in the n/a values in the Pandas dataframe.

      I would recommend using statistics or a model as well and compare results.

      Reply
  11. Amit December 29, 2017 at 5:33 pm #

    Hi Jason,

    I am trying to prepare data for the TITANIC dataset. One of the columns is CABIN which has values like ‘A22′,’B56’ and so on. This column has maximum number of missing values. First I thought to delete this column but I think this could be an important variable for predicting survivors.

    I am trying to find a strategy to fill these null values. Is there a way to fill alphanumeric blank values?

    If there is no automatic way, I was thinking of fill these records based on Name, number of sibling, parent child and class columns. E.g. for a missing value, try to see if there are any relatives and use their cabin number to replace missing value.

    Similar case is for AGE column which is missing. Any thoughts?

    Reply
    • Jason Brownlee December 30, 2017 at 5:19 am #

      Sounds like a categorical variable. You could encode them as integers. You could also assign an “unknown” integer value (e.g. -999) for the missing value.

      Perhaps you can develop a model to predict the cabin number from other details and see if that is skilful.

      Reply
  12. Chidoooo February 9, 2018 at 10:15 pm #

    Good day, I ran this file code pd.read_csv(r’C:\Users\Public\Documents\SP_dow_Hist_stock.csv’,sep=’,’).pct_change(252)
    and it gave me missing values (NAN) of return of stock. How do I resolve it.

    pd.read_csv(r’C:\Users\Public\Documents\SP_dow_Hist_stock.csv’,sep=’,’)
    Out[5]:
    Unnamed: 0 S&P500 Dow Jones
    0 Date close Close
    1 1-Jan-17 2,275.12 24719.22
    2 1-Jan-16 1,918.60 19762.60
    3 1-Jan-15 2,028.18 17425.03
    4 1-Jan-14 1,822.36 17823.07
    5 1-Jan-13 1,480.40 16576.66
    6 1-Jan-12 1,300.58 13104.14
    7 1-Jan-11 1,282.62 12217.56
    8 1-Jan-10 1,123.58 11577.51
    9 1-Jan-09 865.58 10428.05
    10 1-Jan-08 1,378.76 8776.39
    11 1-Jan-07 1,424.16 13264.82
    12 1-Jan-06 1,278.73 12463.15
    13 1-Jan-05 1,181.41 10717.50
    14 1-Jan-04 1,132.52 10783.01
    15 1-Jan-03 895.84 10453.92
    16 1-Jan-02 1,140.21 8341.63
    17 1-Jan-01 1,335.63 10021.57
    18 1-Jan-00 1,425.59 10787.99
    19 1-Jan-99 1,248.77 11497.12
    20 1-Jan-98 963.36 9181.43
    21 1-Jan-97 766.22 7908.25
    22 1-Jan-96 614.42 6448.27
    23 1-Jan-95 465.25 5117.12
    24 1-Jan-94 472.99 3834.44
    25 1-Jan-93 435.23 3754.09
    26 1-Jan-92 416.08 3301.11
    27 1-Jan-91 325.49 3168.83
    28 1-Jan-90 339.97 2633.66
    29 1-Jan-89 285.4 2753.20
    .. … … …
    68 1-Jan-50 16.88 235.42
    69 1-Jan-49 15.36 200.52
    70 1-Jan-48 14.83 177.30
    71 1-Jan-47 15.21 181.16
    72 1-Jan-46 18.02 177.20
    73 1-Jan-45 13.49 192.91
    74 1-Jan-44 11.85 151.93
    75 1-Jan-43 10.09 135.89
    76 1-Jan-42 8.93 119.40
    77 1-Jan-41 10.55 110.96
    78 1-Jan-40 12.3 131.13
    79 1-Jan-39 12.5 149.99
    80 1-Jan-38 11.31 154.36
    81 1-Jan-37 17.59 120.85
    82 1-Jan-36 13.76 179.90
    83 1-Jan-35 9.26 144.13
    84 1-Jan-34 10.54 104.04
    85 1-Jan-33 7.09 98.67
    86 1-Jan-32 8.3 60.26
    87 1-Jan-31 15.98 77.90
    88 1-Jan-30 21.71 164.58
    89 1-Jan-29 24.86 248.48
    90 1-Jan-28 17.53 300.00
    91 1-Jan-27 13.4 200.70
    92 1-Jan-26 12.65 157.20
    93 1-Jan-25 10.58 156.66
    94 1-Jan-24 8.83 120.51
    95 1-Jan-23 8.9 95.52
    96 1-Jan-22 7.3 98.17
    97 1-Jan-21 7.11 80.80

    [98 rows x 3 columns]

    pd.read_csv(r’C:\Users\Public\Documents\SP_dow_Hist_stock.csv’,sep=’,’).pct_change(251)
    Out[7]:
    Unnamed: 0 S&P500 Dow Jones
    0 NaN NaN NaN
    1 NaN NaN NaN
    2 NaN NaN NaN
    3 NaN NaN NaN
    4 NaN NaN NaN
    5 NaN NaN NaN
    6 NaN NaN NaN
    7 NaN NaN NaN
    8 NaN NaN NaN
    9 NaN NaN NaN
    10 NaN NaN NaN
    11 NaN NaN NaN
    12 NaN NaN NaN
    13 NaN NaN NaN
    14 NaN NaN NaN
    15 NaN NaN NaN
    16 NaN NaN NaN
    17 NaN NaN NaN
    18 NaN NaN NaN
    19 NaN NaN NaN
    20 NaN NaN NaN
    21 NaN NaN NaN
    22 NaN NaN NaN
    23 NaN NaN NaN
    24 NaN NaN NaN
    25 NaN NaN NaN
    26 NaN NaN NaN
    27 NaN NaN NaN
    28 NaN NaN NaN
    29 NaN NaN NaN
    .. … … …
    68 NaN NaN NaN
    69 NaN NaN NaN
    70 NaN NaN NaN
    71 NaN NaN NaN
    72 NaN NaN NaN
    73 NaN NaN NaN
    74 NaN NaN NaN
    75 NaN NaN NaN
    76 NaN NaN NaN
    77 NaN NaN NaN
    78 NaN NaN NaN
    79 NaN NaN NaN
    80 NaN NaN NaN
    81 NaN NaN NaN
    82 NaN NaN NaN
    83 NaN NaN NaN
    84 NaN NaN NaN
    85 NaN NaN NaN
    86 NaN NaN NaN
    87 NaN NaN NaN
    88 NaN NaN NaN
    89 NaN NaN NaN
    90 NaN NaN NaN
    91 NaN NaN NaN
    92 NaN NaN NaN
    93 NaN NaN NaN
    94 NaN NaN NaN
    95 NaN NaN NaN
    96 NaN NaN NaN
    97 NaN NaN NaN

    [98 rows x 3 columns]

    Reply
    • Jason Brownlee February 10, 2018 at 8:57 am #

      Perhaps post your code and issue to stackoverflow?

      Reply
  13. Ravi March 13, 2018 at 10:59 pm #

    Hi Jason,

    Thanks for your valuable writing.
    I have one question :-
    We can also replace NaN values with Pandas fillna() function. In my opinion this is more versatile than Imputer class because in a single statement we can take different strategies on different column.
    df.fillna({‘A’:df[‘A’].mean(),’B’:0,’C’:df[‘C’].min(),’D’:3})

    What is your opinion? Is there any performance difference between two?

    Reply
    • Jason Brownlee March 14, 2018 at 6:23 am #

      Great tip.

      No idea, on the performance difference.

      Reply
  14. annusin0_0 March 26, 2018 at 4:31 am #

    Is there a recommended ratio on the number of NaN values to valid values , when any corrective action like imputing can be taken?
    If we have a column with most of the values as null, then it would be better off to ignore that column altogether for feature?

    Reply
    • Jason Brownlee March 26, 2018 at 10:04 am #

      No, it is problem specific. Perhaps run some experiments to see how sensitive the model is to missing values.

      Reply
  15. Ammar Hasan March 31, 2018 at 2:35 pm #

    Hi Jason,

    Thanks for this post, I wanted to ask, how do we impute missing text values in a column which has either text labels or blanks.

    Reply
    • Jason Brownlee April 1, 2018 at 5:44 am #

      Good question, I’m not sure off hand. Perhaps start with simple masking of missing values.

      Reply
    • rakend dubba May 27, 2018 at 6:05 pm #

      To fill the nan for a categorical column

      df = df.fillna(df[‘column’].value_counts().index[0])
      This fills the missing values in all columns with the most frequent categorical value

      Reply
  16. Gabriel April 6, 2018 at 9:44 pm #

    Thanks a lot Jason ! but I have a little question, how about if we want to replace missing values with the mean of each ROW not column ? how to do that ? if you have any clue, please tell me.. Thank you again Jason.

    Reply
    • Jason Brownlee April 7, 2018 at 6:32 am #

      Why would you do this?

      numpy.mean() allows you to specify the axis on which to calculate the mean. It will do it for you.

      Reply
  17. Adil April 13, 2018 at 8:14 am #

    Hi Jason,

    I wanted to ask you how you would deal with missing timestamps (date-time values), which are one set of predictor variables in a classification problem. Would you flag and mark them as missing or impute them as the mode of the rest of the timestamps?

    Reply
  18. quan April 20, 2018 at 9:19 pm #

    Hi Jason,

    A big fan of yours.

    I have a question about imputing missing numerical values. I don’t really want to remove them and I want to impute them to a value that is like Nan but a numerical type? Would say coding it to -1 work? (0 is already being used).

    I guess I am trying to achieve the same thing as categorising an nan category variable to unknown and creating another feature column to indicate that it is missing.

    Thanks,

    Reply
    • Jason Brownlee April 21, 2018 at 6:49 am #

      NaN is a numerical type. It is a valid float.

      You could use -999 or whatever you like.

      Be careful that your model can support them, or normalize values prior to modeling.

      Reply
  19. Ravi July 12, 2018 at 1:29 pm #

    Hello Jason,

    You mentioned this here: “if you choose to impute with mean column values, these mean column values will need to be stored to file for later use on new data that has missing values.”, but I wanted to ask:

    Would imputing the data before creating the training and test set (with the data set’s mean) cause data leakage? What would be the best approach to tackle missing data within the data pipeline for a machine learning project.

    Let’s say I’m imputing by filling in with the mean. For the model tuning am I imputing values in the test set with the training set’s mean?

    Reply
  20. Tobias August 8, 2018 at 7:36 pm #

    Hi Jason,

    Thanks again for that huge nice article!

    is there a neat way to clean away all those rows that happen to be filled with text (i.e. strings) in a certain column, i.e. List.ImportantColumn .

    This destroys my plotting with “could not convert string to float”

    Thanks already in advance!

    Reply
    • Jason Brownlee August 9, 2018 at 7:38 am #

      Yes, you can remove or replace those values with simple NumPy array indexing.

      For example, if you have ‘?’ you can do:

      Reply
  21. Anantha Krishnan S September 15, 2018 at 5:45 pm #

    Hi Jason,

    I tried using this dropna to delete the entire row that has missing values in my dataset and after which the isnull().sum() on the dataset also showed zero null values. But the problem arises when i run an algorithm and i am getting an error.

    Error : Input contains NaN, infinity or a value too large for dtype(‘float64’)

    This clearly shows there still exists some null values.

    How do i proceed with this thanks in advance

    Reply
    • Jason Brownlee September 16, 2018 at 5:57 am #

      Perhaps print the contents of the prepared data to confirm that the nans were indeed removed?

      Reply
  22. fatma October 27, 2018 at 2:23 am #

    Hi Jason,

    Thanks for this post, I’m using CNN for regression and after data normalization I found some NaN values on training samples. How can I use imputer to fill missing values in the data after normalization.

    Reply
    • Jason Brownlee October 27, 2018 at 6:03 am #

      Does the above tutorial not help?

      Reply
      • fatma October 29, 2018 at 4:19 pm #

        should I apply Imputer function for both training and testing dataset?

        Reply
        • Jason Brownlee October 30, 2018 at 5:52 am #

          Yes, but if the imputer has to learn/estimate, it should be developed from the training data and aplied to the train and test sets, in order to avoid data leakage.

          Reply
          • fatma October 30, 2018 at 6:24 pm #

            I feel that Imputer remove the Nan values and doesn’t replace them. For example the vector features length in my case is 14 and there are 2 Nan values after applying Imputer function the vector length is 12. This means the 2 Nan values are removed. However I used the following setting:
            imputer = Imputer(missing_values=np.nan, strategy=’mean’, axis=0)

          • Jason Brownlee October 31, 2018 at 6:22 am #

            I don’t know what is happening in your case, perhaps post/search on stackoverflow?

          • fatma November 14, 2018 at 5:22 pm #

            You mean I should fit it on training data then applied to the train and test sets as follow :

            imputer = Imputer(strategy=”mean”, axis=0)
            imputer.fit(X_train)
            X_train = imputer.transform(X_train)
            X_test = imputer.transform(X_test)

          • Jason Brownlee November 15, 2018 at 5:27 am #

            Looks good.

  23. Manik Chand October 28, 2018 at 8:58 pm #

    Thanks for this post!!!
    A dataSet having more than 4000 rows and rows can be groupby their 1st columns and let there is many columns (assume 20 columns) and few columns(let 14 columns) contains NaN(missing value).
    How we populate NaN with mean of their corresponding columns by iterative method(using groupby, transform and apply) .

    Reply
    • Jason Brownlee October 29, 2018 at 5:56 am #

      Sorry, I don’t understand. Perhaps you can elaborate your question?

      Reply
      • Manik Chand October 30, 2018 at 3:19 am #

        actually i want to fill missing value in each column. Value is the mean of corresponding column. Is there any iterative method?

        Reply
        • Jason Brownlee October 30, 2018 at 6:09 am #

          What do you mean by iterative method?

          Reply
          • shailaja March 11, 2020 at 2:57 pm #

            Is it iterative imputer? where missing value acts as dependent variable and independent variables are other features

          • Jason Brownlee March 12, 2020 at 8:36 am #

            No.

  24. Sumod December 27, 2018 at 2:47 pm #

    After replacing zeroes,Can I save it as a new data set?

    Reply
  25. Amit February 7, 2019 at 6:06 pm #

    import numpy as np
    import pandas as pd

    mydata = pd.read_csv(‘diabetes.csv’,header=None)
    mydata.head(20)

    0 1 2 3 4 5 6 7 8
    0 Pregnancies Glucose BloodPressure SkinThickness Insulin BMI DiabetesPedigreeFunction Age Outcome
    1 6 148 72 35 0 33.6 0.627 50 1
    2 1 85 66 29 0 26.6 0.351 31 0
    3 8 183 64 0 0 23.3 0.672 32 1
    4 1 89 66 23 94 28.1 0.167 21 0
    5 0 137 40 35 168 43.1 2.288 33 1

    print((mydata[0] == 0).sum()) — for any column it always shows 0
    0 >>>>>>>…. any thing wrong here ?

    whereas i have 0’s in dataset

    0 Pregnancies
    1 6
    2 1
    3 8
    4 1
    5 0>>>>>>>>>
    6 5
    7 3
    8 10
    9 2
    10 8
    11 4
    12 10
    13 10
    14 1
    15 5
    16 7
    17 0 >>>>>>

    Reply
    • Jason Brownlee February 8, 2019 at 7:45 am #

      Perhaps post your code and issue to stackoverflow?

      Reply
    • Pauline May 14, 2020 at 8:49 pm #

      Hello,

      More than one year later, I have the same problem as you. When i search for 0 it does not work. However, when I look for ‘0’ it does, which means the table is filled with strings and not number… Any idea how I can handle that?

      Best Regards

      Reply
      • Jason Brownlee May 15, 2020 at 6:00 am #

        Perhaps your data was loaded as strings?

        Try converting it to numbers:

        Reply
  26. Krishna March 24, 2019 at 3:59 am #

    Hi sir,
    For my data after executing following instructions still I get same error
    dataset= dataset.replace(0, np.NaN)
    dataset.dropna(inplace=True)
    dataset= dataset.replace(0, np.Inf)
    dataset.dropna(inplace=True)
    print(dataset.describe())
    F1 F2 F3 F4
    count 1200.000000 1200.000000 1200.000000 1200.000000
    mean 0.653527 0.649447 1.751579 inf
    std 0.196748 0.194933 0.279228 NaN
    min 0.179076 0.179076 0.731698 0.499815
    25% 0.507860 0.506533 1.573212 1.694007
    50% 0.652066 0.630657 1.763520 1.925291
    75% 0.787908 0.762665 1.934603 2.216663
    max 1.339335 1.371362 2.650390 inf

    How can I get out from this problem.

    Reply
    • Jason Brownlee March 24, 2019 at 7:07 am #

      Sorry to hear that, perhaps try posting your code and question to stackoverflow?

      Reply
    • Raj January 9, 2020 at 11:27 pm #

      df.replace(-np.Inf, 0 )
      df.replace(np.Inf, 0 )

      Reply
  27. mubasher March 28, 2019 at 6:38 pm #

    how can we impute the categorical data in python

    Reply
    • Jason Brownlee March 29, 2019 at 8:29 am #

      You can use an integer encoding (label encoding), a one hot encoding or even a word embedding.

      Reply
  28. Rotenda Takalani April 16, 2019 at 9:42 pm #

    Hi Jason

    I am new to Python and I was working through the example you gave. For some reason, When I run the piece of code to count the zeros, the code returns results that indicate that there are no zeros in any of those columns.

    Can you please assist?

    Reply
  29. chu June 4, 2019 at 7:20 pm #

    Hi Jason,

    Great post. Thanks so much.

    Say I have a dataset without headers to identify the columns, how can I handle inconsistent data, for example, age having a value 2500 without knowing this column captures age, any thoughts?

    Reply
  30. Muhammad Irfan November 26, 2019 at 7:02 am #

    Hi Jason,

    Nice article. How can we add (python) another feature indicating a missing value as 1 if available and 0 if not? Is that a sensible solution?

    Thank you.

    Reply
    • Jason Brownlee November 26, 2019 at 1:28 pm #

      Great question.

      You could loop over all rows and mark 0 and 1 values in a another array, then hstack that with the original feature/rows.

      Reply
  31. Hadis December 11, 2019 at 1:44 pm #

    Pima Indians Diabetes Dataset doesn’t exist anymore 🙁

    Reply
  32. Hussain December 12, 2019 at 9:26 pm #

    What is the current situation in AutoML field? What researchers try to bring out actually?

    Reply
    • Jason Brownlee December 13, 2019 at 6:00 am #

      Good question, I need to learn more about that field.

      Reply
  33. Hussain December 13, 2019 at 9:21 pm #

    Let me know ,once you get to know about that someday.
    Thank you for your response!!

    Reply
  34. Bill January 1, 2020 at 3:17 pm #

    [Ignore earlier misplaced post]

    Jason,
    Many thanks for your work in preparing these awesome tutorials!

    In order to fill missing values with mean column values, I had to switch from:
    from sklearn.preprocessing import Imputer
    …
    # fill missing values with mean column values
    imputer = Imputer()

    To:
    from sklearn.impute import SimpleImputer
    …
    imputer = SimpleImputer(missing_values=numpy.NaN, strategy=’mean’)

    Reply
  35. German Niebles January 5, 2020 at 4:00 am #

    Jason, thanks a lot for your article,very useful.
    Regards

    Reply
  36. Raj January 30, 2020 at 6:12 pm #

    Hi Jason, great tutorial! If I were to impute values for time series data, how would I need to approach it? My dataset has data for a year and data is missing for about 3 months. Is there any way to salvage this time series for forecasting?

    Reply
  37. Shreya February 29, 2020 at 3:37 am #

    I have near about 4 lakhs of data.
    The shape of my dataset is (400000,114).
    I want to first impute the data and then apply feature selection such as RFE so that I could train my model with only the important features further instead of all 114 features.
    But I am unable to understand how after using SimpleImputer and MinMax scaler to normalize the data as :

    values = dataset.values
    imputer = SimpleImputer()
    imputedData = imputer.fit_transform(values)
    scaler = MinMaxScaler(feature_range=(0, 1))
    normalizedData = scaler.fit_transform(imputedData)

    How will we use this normalized data ??
    Because on normal dataset further I am making X,Y labels as:

    X = dataset.drop([‘target’], axis=1)
    y = dataset.target

    How RFE will be used here further ?
    Whether on X and y labels or before that do we have to convert all X labels to normalized data ?

    Reply
    • Shreya February 29, 2020 at 3:41 am #

      Also training this huge amount of data with Random Forest or Logistic Regression for RFE is taking much of time ? So is a better solution available for training ?

      Reply
      • Jason Brownlee February 29, 2020 at 7:20 am #

        Perhaps use a smaller sample of your data to start with.

        Reply
        • Shreya February 29, 2020 at 3:07 pm #

          I have tried it with smaller set of data which is working fine.
          But in a requirement I have to use this large sized i.e. 4 lakhs of data with 114 features.
          Also RFE on RandomForest is taking a huge amount of time to run.
          And if I go with model = LogisticRegression(‘saga’), then the amount of time is less but I am dealing with warnings which I am unable to resolve as:

          ConvergenceWarning: The max_iter was reached which means the coef_ did not converge
          “the coef_ did not converge”, ConvergenceWarning)

          How should I go further for feature selection on this large dataset ?

          Thank you very much !!

          Reply
          • Jason Brownlee March 1, 2020 at 5:22 am #

            Perhaps fit on less data, at least initially.

    • Jason Brownlee February 29, 2020 at 7:20 am #

      I would recommend developing a pipeline so that the imputation can be applied prior to scaling and feature selection and the prior to any modeling.

      Reply
  38. manjunath February 29, 2020 at 6:00 am #

    Hi Jason I have Time Series Data so i need to fill missing values , so which is best technique to fill time series data ?

    Reply
  39. Bruno Campetella April 4, 2020 at 2:05 am #

    Hello Jason
    If we impute a column in our dataset the data distribution will change, and the change will depend on the imputation strategy. This in turns will affect the different ML algorithms performance. We are tuning the prediction not for our original problem but for the “new” dataset, which most probably differ from the real one. My question is, for avoiding error predictions or overestimated performance of our algorithm, shouldn’t we avoid having any NA’s imputed values in our test dataset? I guess We can use them in the training dataset and using different imputation techniques to check performance of the algorithms on the test data (without imputed NA’s).
    Thanks
    Bruno

    Reply
    • Jason Brownlee April 4, 2020 at 6:25 am #

      Yes. Great question!

      Test a few strategies and use the approach that results in a model that has the best skill.

      Reply
  40. Deeksha Mahapatra May 11, 2020 at 3:42 am #

    Hi Jason,
    First of all great job on the tutorials! This is my go to place for Machinel earning now.

    I am trying to impute values in my dataset conditionally. Say I have three columns, If Column 1 is 1 then Column 2 is 0 and Column 3 is 0; If column 1 is 2 then Column 2 is Mean () and Column 3 is Mean(). I tried running an if statement with the function any() and defined the conditions separately. However the conditions are not being fulfilled based on conditions, I am either getting all mean values or all zeroes.
    I have posted this on Stackoverflow and haven’t gotten any response to help me with this.Please do suggest what should I apply to get this sorted.

    Thanks a lot!

    Reply
    • Jason Brownlee May 11, 2020 at 6:08 am #

      Thanks!

      Perhaps try writing the conditions explicitly and enumerate the data, rather than using numpy tricks? It will be slower but perhaps easier to debug.

      Reply
      • Deeksha Mahapatra May 11, 2020 at 4:43 pm #

        Yes, I used iloc to define the conditions separately. Worked fine. Thanks!

        Reply
  41. Parthiv June 14, 2020 at 8:27 am #

    Thanks for the post.

    Please correct me if i am wrong.

    Applying these techniques for training data works for me. However, if the data in real-time (test data) is received with standard inverval (100 milliseconds), then algorithms suchs as LGBM, XGBoost and Catboost (scikit) with inherent capabilities can be used.

    Your Weka post on missing values by defining threshold works great.

    Reply
    • Jason Brownlee June 15, 2020 at 5:58 am #

      100ms is a long time for a computer, I don’t see the problem with using imputation.

      Reply
      • Parthiv June 19, 2020 at 6:58 am #

        Thanks for the reply.

        Just a clarification. If one instance of data from several sensors arrive with some missing values for every 100ms, is it possible to classify based on the current instance alone. (one instance at a time).

        My presumption is that we need multiple instances to calculate the statistics even for stream data.

        Sorry. A bit confused on this.

        Reply
        • Jason Brownlee June 19, 2020 at 1:10 pm #

          Generally, you can frame the prediction problem any way you wish, e.g. based on the data you have and the data you need at prediction time.

          Then train a model based on that framing of the problem.

          Reply
          • Parthiv June 19, 2020 at 3:27 pm #

            Thanks a lot replying with patience.

          • Jason Brownlee June 20, 2020 at 6:06 am #

            I’m here to help if I can.

  42. Anthony The Koala July 5, 2020 at 10:31 pm #

    Dear Dr Jason,
    Background information and question:

    Background information:
    Instead of playing around with the “horse colic” data with missing data, I constructed a smaller version of the iris data. I had to shuffle the data to get an even spread of species 0, 1 or 2. Otherwise if I took the first 20 rows the last column would be full of species 0. Hence my shuffling of the data.

    I’ve had great success in predicting the kind of species.

    So my iris20 data looks like this – the first four columns are in the correct order of the original iris data and the last column are a variety of species. .

    I removed 10 values ‘at random’ from my iris20 data, called it iris20missing

    Question:
    I have successfully been able to predict the kind of species of iris whether it is species 0, 1, 2.
    Examples:

    My question: In listing 8.19, 3rd last line, page 84 (101 of 398):

    row is enclosed in brackets [row].
    that is we have for example row = [[6.3 ,NaN,4.4 ,1.3]]
    Why please do we double enclose the array in predict function?

    When I do this

    I get errors.

    Thank you for your time,
    Anthony of Sydney

    Why enclose row as [row] since row is already enclosed by brackets.
    That is why .predict([row]) and not .predict(row)

    Reply
    • Jason Brownlee July 6, 2020 at 6:35 am #

      Nice work!

      The predict() function expects a 2d matrix input, one row of data represented as a matrix is [[a,b,c]] in python.

      Reply
      • Anthony The Koala July 6, 2020 at 9:35 am #

        First,
        Thanks in advance for your reply. It is appreciated.

        A question on your answer please.

        Background info

        In the above example we had to structure the variable ‘row’ as a 2d matrix for use in the predict() function. Here ‘row’ is changed from an array of size 4 to a 1 x 4 matrix.

        I’ve worked out that one can construct an n x m matrix and have the model predict for an n x m matrix

        To illustrate:

        Recall in my above example I made a series of rows and made individual predictions on the model with these rows:

        Now if we made an n x m matrix and feed that n x m matrix into the predict() function we should expect the same outcomes as individual predictions.

        Result is the same as if making individual predictions. Hence I understand the predict() function expecting a matrix and if predicting for single rows, make the single row into a 1xm matrix.

        Conclusion: the predict() function expects a matrix, and we can make an n x m matrix containing the rows of what we want to predict AND get multiple results.

        Thank you again in advance
        Anthony of Sydney

        Reply
        • Jason Brownlee July 6, 2020 at 2:06 pm #

          Yes.

          Perhaps this will help clarify:
          https://machinelearningmastery.com/make-predictions-scikit-learn/

          Reply
          • Anthony The Koala July 6, 2020 at 7:29 pm #

            Dear Dr Jason,
            Thank you for the blog at https://machinelearningmastery.com/make-predictions-scikit-learn/.

            Relevant to answer my question about prediction are the sections “Class Predictions”, “Single Class Predictions” and “Multiple Class Predictions”. (These are presented in order of 1, 3 and 2 ).

            The variable Xnew is of the structure [[],[]] which is a 2D structure. This is for one prediction.

            In the multiple class predictions, Xnew is a 2D matrix. In the example it is a 3 x 2 2D matrix

            In both cases of single or multiple class predictions we feed the 2D matrix in the form

            In sum predicting requires our feature matrix to be 2D whether 1 x m or n x m, where 1 or n are the number of predictions and m being the number of features.

            Thank you,
            Anthony of Sydney

  43. Anthony The Koala July 14, 2020 at 10:02 pm #

    Dear Dr Jason,
    I wish to share my two ways of deleting specific rows from a dataset as per subheading 4, “Remove Rows With Missing Values”

    HOW TO DELETE SPECIFIC VALUES FROM SPECIFIC COLUMNS – TWO METHODS
    Method #1 as per heading 4 = listing 7.16 on p73 (90 of 398) of your book.

    Method #2 – using arrays

    The last method was presented in case your data set is not as a DataFrame.

    Thank you,
    Anthony of Sydney

    Reply

Leave a Reply