Addressing Absence: Handling Missing Values

Table of Contents

I. Introduction

Definition of Missing Values

In the world of data, we often come across tables or datasets with empty spaces. These empty spaces, where information is absent or missing, are referred to as Missing Values. Think of it like this, you have a box of your favorite cookies. When you open it, you find some spaces where cookies should be, but they’re not there. Those empty spots are similar to missing values in data.

Brief Explanation of Missing Values in Datasets

Now, let’s imagine a class of students who took a test. The scores of all students are recorded in a table. However, some students were absent for the test, and hence their scores are not recorded in the table. In the dataset of test scores, the absent students’ scores are missing values.

Data can be missing for a variety of reasons. It might be that the information was never recorded, got lost, or is not applicable in certain cases. Just like in our class example, the scores for the absent students were never recorded because they didn’t take the test. Hence, the scores for these students are missing values in our dataset.

Missing values are like puzzles in our data. They can be tricky to deal with because we don’t always know why the information is missing or what the missing information might have been.

Importance of Handling Missing Values in Machine Learning

If we ignore these missing values and try to create our machine-learning models, it can lead to big problems. Our models might make incorrect predictions or might not work at all because they don’t know how to deal with the missing information.

Therefore, it’s very important to handle these missing values before we use our data to build machine-learning models. This process is part of a step in data analysis called Data Cleaning or Data Preprocessing. Think of it like cleaning your room. You have to sort and arrange everything before you can easily find what you’re looking for.

In the next sections, we’ll learn more about missing values, why they are important, and how we can handle them in different ways. We’ll use simple language, fun examples, and even some Python code to make this journey interesting for you. So, let’s get started!

II. Theoretical Understanding of Missing Values

Let’s jump into the ocean of missing values. Don’t worry, we have our life jackets on! So, let’s dive deeper.

Concept and Types of Missing Values: MCAR, MAR, MNAR

Just like different types of cookies have different flavors, missing values also have different types. But unlike cookies, we usually have three types of missing values. We give them funny-sounding names: MCAR, MAR, and MNAR.

  1. MCAR stands for Missing Completely At Random. This means that the values are missing randomly, and it has nothing to do with any other data in your dataset. It’s like a lottery system, where every piece of data has an equal chance of going missing. For example, suppose you have a list of student names and their scores. If a few scores are missing and it has nothing to do with their names or any other data you have about these students, then these missing values are MCAR.
  2. MAR stands for Missing At Random. This is when the missingness of data has something to do with some of your other data, but not the missing data itself. Let’s go back to our student list example. Suppose the students’ scores are missing, and you notice that most of the missing scores are from students who are absent a lot. Here, the missing scores (missing data) are related to the students’ absences (other data), but not to the scores themselves. These missing values are MAR.
  3. MNAR stands for Missing Not At Random. This is when the missingness of data has a relationship with the missing data itself. That sounds confusing, doesn’t it? Let’s explain with an example. Imagine you have a survey where people can choose not to answer certain questions. If a question like “Do you smoke?” has many missing answers, it’s possible that people who smoke didn’t want to answer this question. Here, the missingness (missing data) is related to the answer to the question (the missing data itself). These missing values are MNAR.

Each type of missing data requires different handling methods, which we’ll learn more about later. For now, let’s remember these funny-sounding names and their meanings.

Mathematical Foundation: The Impact on Statistical Analysis

You might be wondering, why are missing values a problem? It’s because they can mess up our calculations and results. Just like missing ingredients can ruin a recipe, missing values can distort our statistical results.

  1. Average (Mean): Suppose you’re calculating the average score of a class, but some scores are missing. If you calculate the average with the scores you have, you might get a higher or lower average than the real one, because the missing scores are not included.
  2. Variance: Variance tells us how spread out our data is. Missing values can affect this too. If the missing scores were very high or very low, they could change how spread out the scores are.
  3. Correlation: Correlation tells us how related two sets of data are. If some data is missing, it can change how related we think our data is.

So, missing values can play a big trick on us by giving us incorrect results. This is why it’s important to handle them correctly.

Understanding the Nature of Missingness

Before we decide how to handle our missing values, we need to understand their nature. Are they MCAR, MAR, or MNAR? Each type requires a different method of handling.

To understand the nature of missingness, we need to be a bit like detectives. We have to look at our data, think about where it came from, and look for patterns. If the missing data seems to be random and not related to any other data, it might be MCAR. If it’s related to other data but not to itself, it might be MAR. If it’s related to itself, it’s probably MNAR.

As we’ve seen, understanding missing values isn’t just about spotting the empty spaces in our data. It’s also about understanding why they’re empty and what it means for our analysis. With this understanding, we can choose the best way to handle our missing values. But more on that later.

III. Consequences of Ignoring Missing Values

Data is like the ingredients for our machine-learning recipe. If some ingredients are missing and we ignore them, our recipe might not turn out as we expected. Similarly, if we ignore missing values in our data, it can lead to some big problems. Let’s understand these problems one by one:

1. Distortion of Statistical Results

Remember how missing values can mess up our calculations of mean, variance, and correlation? If we ignore these missing values, we might get incorrect results, and our analysis might be off. It’s like trying to bake a cake without sugar and expecting it to be sweet!

Let’s consider an example where we have scores of 10 students and two of them didn’t take the test. The scores of the remaining 8 students are 6, 7, 8, 7, 7, 8, 9, and 8. If we calculate the average without considering the missing scores, we get an average score of 7.5. But what if the missing scores were really low or really high? That would change our average, right?


So, by ignoring missing values, we’re potentially distorting our statistical results.

2. Inaccurate Machine Learning Models

In machine learning, we feed data to our model, and the model learns patterns from the data to make predictions or decisions. If our data has missing values and we ignore them, our model might not learn correctly.

Let’s imagine that our model is a student and our data is the study material. If some topics (missing values) are missing from the study material, the student (model) might not perform well in the exam (making predictions or decisions), right?

So, ignoring missing values can make our machine-learning models inaccurate.

3. Influence on Data Distribution and Variance

In statistics, distribution is a way that values in a dataset spread out, and variance is how much these values vary. Missing values can influence both distribution and variance.

Imagine you’re making a necklace with different kinds of beads. The colors and shapes of the beads determine the look of your necklace (distribution). If some beads (values) are missing, your necklace might look different.

Similarly, if you have a class of students with different heights, and some students (values) are missing, it might change the average height (mean) and how different each student’s height is from the average (variance).

So, ignoring missing values can affect the distribution and variance of our data, which in turn can affect our analysis and machine learning models.

Distortion of Statistical ResultsIncorrect calculations and analysisIncorrect average of student scores
Inaccurate Machine Learning ModelsPoor performance of modelsModel fails to make accurate predictions
Influence on Data Distribution and VarianceChanges the data structureChanges the look of the necklace or the average height of students

IV. Basic Techniques of Handling Missing Values

When we encounter missing values in our dataset, it’s like we are trying to complete a jigsaw puzzle with missing pieces. It’s hard to see the full picture without all the pieces. Luckily, there are several basic techniques that we can use to deal with missing values. Let’s understand them one by one:

1. Listwise and Pairwise Deletion

Listwise deletion, also known as complete-case analysis, is like playing a game of football with full teams. If a team is missing a player (a row in our dataset is missing a value), we simply don’t consider that match (we remove that row from our dataset).

For example, if we have a dataset of students’ test scores and two students didn’t take the test, we simply remove those students from our analysis.

Pairwise deletion, on the other hand, is a little more flexible. It’s like saying, “If a player is missing for one match, we can still play the other matches with that team.” In other words, we only exclude the missing values when we actually need them for a specific analysis, but keep them for other analyses.

Here’s a table to help you understand:

Listwise DeletionRemoves the entire row if a single value is missing.Remove a student’s record if the test score is missing.
Pairwise DeletionRemoves the missing values for specific analyses where they are required.Exclude missing test scores only when calculating the average score, but include them for other analyses like finding the number of students.

2. Mean, Median, and Mode Imputation

Mean imputation is like filling up a hole in the road with cement. If a value is missing (there’s a hole), we fill it up with the average value (cement).

For instance, if we are missing a student’s test score, we can fill it in with the average score of all the students. Similarly, we can use the median (middle value) or mode (most frequently occurring value) to fill in the missing values.

Here’s how it works:

Mean ImputationFills missing values with the mean (average) value of the other entries.Fill a missing test score with the average score of all students.
Median ImputationFills missing values with the median (middle) value of the other entries.Fill a missing test score with the median score of all students.
Mode ImputationFills missing values with the mode (most common) value of the other entries.Fill a missing test score with the most common score of all students.

3. Random Sampling Imputation

Random sampling imputation is like drawing a lucky draw. If a value is missing, we randomly pick a value from the rest of the data. It’s a little like saying, “If a student didn’t take a test, let’s assume they would get a score similar to one of the other students.”

This table summarizes it:

Random Sampling ImputationFills missing values with a randomly selected value from the other entries.Fill a missing test score with a score randomly selected from the other students’ scores.

It’s important to remember that these basic techniques are simple and easy to use, but they may not always give the best results. It’s like trying to fix a car with only a basic toolkit. You can do some repairs, but for some problems, you might need more advanced tools. In the next section, we will explore some advanced techniques for handling missing values.

V. Advanced Techniques of Handling Missing Values

When you want to assemble a complicated puzzle, sometimes you need more than just the basic tools. Similarly, in dealing with missing values, basic techniques might not always be enough. So, let’s take out our advanced toolkit and see how we can use more complex methods to fill in those blanks!

1. Regression Imputation

Have you ever tried to guess a friend’s score in a game based on how they usually play? Regression imputation is something like that. It’s like being a detective and using clues (other variables) in our data to guess the missing value.

For example, if we’re missing a student’s math score, but we know their scores in physics and chemistry, we can use a mathematical relation (regression) between these subjects to predict the missing math score.

Here’s how it works:

Regression ImputationUses a regression model to predict and fill missing values based on other variables.Predict a missing math score using known physics and chemistry scores.

But remember, like guessing your friend’s score, this method can sometimes guess wrong, especially if there is no clear relation between the variables.

2. K-Nearest Neighbors (KNN) Imputation

Imagine you’re in a fruit market trying to guess the weight of an apple. One way to do it could be by finding a few apples that are similar in size and shape to your apple and then guessing your apple’s weight based on them. This is similar to how KNN imputation works.

In KNN imputation, we find a few data points (neighbors) that are similar to the data point with the missing value, and then we guess the missing value based on these neighbors.

For instance, if we want to guess a student’s math score, we find other students who have similar scores in other subjects and then use their math scores to guess the missing score.

Let’s summarize it:

K-Nearest Neighbors (KNN) ImputationUses the values of the nearest data points (neighbors) to fill missing values.Predict a missing math score using the math scores of students with similar scores in other subjects.

While this method can be quite effective, just like in our apple example, if our apples are not very similar, our guess might not be very accurate.

3. Multiple Imputation

Did you ever fill out a multi-choice quiz by making several educated guesses and then choosing the most common or average answer? That’s sort of what multiple imputation is about.

Multiple imputation involves making multiple guesses (imputations) for each missing value, creating several different complete datasets. Each of these datasets is analyzed separately, and the results are pooled to give a final answer.

For example, if we’re missing a student’s math score, we make several educated guesses for it, analyze our data with each guess, and then combine the results.

Here’s the idea:

Multiple ImputationMakes multiple guesses for each missing value and combines the results.Make several guesses for a missing math score, analyze the data with each guess, and combine the results.

This method is like having multiple shots at a target. It can give more accurate results, but it also takes more time and effort.

As you can see, these advanced techniques provide more sophisticated ways to handle missing values. However, just like advanced tools, they require more understanding and careful handling. In the next section, we’ll compare these techniques and help you choose the right one for your data puzzle.

Remember: the best technique depends on your data and your goal. It’s like choosing the right tool for the job. You wouldn’t use a hammer to fix a glass vase, would you?

VI. Comparing Missing Values Handling Techniques

Have you ever been to a store and seen several items that seem to do the same job, and you wondered, “Which one is the best?” This is a common situation when we have several methods to handle missing values in our data. We have a whole toolset, from simple to advanced, and each tool has its own strengths and weaknesses.

So, how do we decide which tool to use? Let’s compare these tools (methods) to understand their strengths and weaknesses. This will help us choose the right one for our job. Don’t worry! I’ve organized this information in easy-to-understand tables. Let’s dive in!

Comparison with Listwise Deletion

Listwise DeletionIt’s simple and quick. Doesn’t distort correlations among observed values.It can waste valuable information when the missingness is random. Risk of bias if data is not missing completely at random (MCAR).

Comparison with Pairwise Deletion

Pairwise DeletionIt retains more data compared to listwise deletion. Good for handling MCAR and MAR data.Results can be inconsistent due to different numbers of observations being used for different analyses.

Comparison with Mean, Median, and Mode Imputation

Mean ImputationIt’s simple and maintains the overall data mean.It reduces data variability and weakens correlations with other variables. Not suitable for non-normal data.
Median ImputationRobust to outliers. Maintains the overall data median.Similar to mean imputation, it reduces data variability. Not suitable for non-normal data.
Mode ImputationEasy to implement. Suitable for categorical data.Can introduce bias if the missingness is related to the variable.

Comparison with Regression Imputation

Regression ImputationIt uses correlations with other variables. More accurate than mean, median, and mode imputation.It assumes a perfect relationship between variables, which is rarely the case. Can underestimate data variability.

Comparison with K-Nearest Neighbors (KNN) Imputation

KNN ImputationIt takes into account the variable correlation structure. It works well with multivariate data.It can be computationally intensive for large datasets. It assumes that the dataset has similar cases.

Comparison with Multiple Imputation

Multiple ImputationIt accounts for uncertainty in the imputation process. It’s flexible and can handle different variable types.It can be complex and computationally intensive. It requires expertise to implement correctly.

Now, you might be thinking, “That’s a lot of information! How do I remember all of this?” Well, don’t worry! We have explored these techniques in detail in our earlier sections. Just remember the golden rule: The best technique depends on your data and your goal. It’s like choosing the right tool for the job.

Choosing the right tool is not always about picking the most advanced or sophisticated one. It’s about understanding the job at hand and the material we’re working with. After all, you wouldn’t use a bulldozer to pluck a flower, would you? So, always start by understanding your data!

VII. Missing Values Handling in Action: Practical Implementation

Alright! Now that we have a strong understanding of the various techniques to handle missing values, let’s put them to the test with some hands-on practice. The best way to learn is by doing, right? We’re going to apply these techniques to a real-world dataset and see how they work in action.

To keep things simple and accessible, we’ll be using the Iris dataset from the scikit-learn library. The Iris dataset is widely used for data science and machine learning tutorials, mainly due to its simplicity and the variety of patterns it holds.

Choosing a Dataset

The Iris dataset is a perfect choice for our task because it’s relatively simple, but also complex enough to illustrate the challenges of dealing with missing data. It includes measurements of sepal length, sepal width, petal length, and petal width for three different species of Iris flowers.

However, the original Iris dataset doesn’t have any missing values. To create a more realistic scenario, we’re going to manually introduce some missing values into the dataset. But don’t worry, we’ll keep track of what we’re doing so we can evaluate our techniques later.

Data Exploration and Visualization

Before diving into handling missing values, let’s have a quick look at our dataset. The describe() function in pandas gives us a summary of the numerical attributes, such as count, mean, min, and max.


However, it’s often more intuitive to understand data by visualizing it. Histograms are a great way to visualize the distribution of a dataset. We’ll use the matplotlib and seaborn libraries to draw these histograms. Don’t forget, always visualizing your data before and after treating missing values is an essential step to understand the effect of the treatment.

Data Preprocessing: Identifying Missing Values

We’ve introduced some missing values into the dataset randomly. Now, it’s time to identify these missing values. The isnull().sum() function from pandas helps us to see the number of missing values in each column.

print(f"Initial Missing Values:\n{df.isnull().sum()}\n")

This will print the number of missing values in each column. Now that we’ve identified the missing values, let’s move on to handling them!

Handling Missing Values with Python Code Explanation

We’re going to use different techniques to handle the missing values: mean, median, and most frequent imputation. We use the SimpleImputer function from the scikit-learn library for this purpose.

We’ve created a dictionary where each key is the name of a technique and the value is an instance of SimpleImputer with the corresponding strategy.

techniques = {
    'Mean': SimpleImputer(strategy='mean'), 
    'Median': SimpleImputer(strategy='median'),
    'Most Frequent': SimpleImputer(strategy='most_frequent'),
    'KNN': KNNImputer(n_neighbors=3),  # Replaces missing values using the mean of 3 nearest neighbors
    'Iterative': IterativeImputer(estimator=ExtraTreesRegressor(n_estimators=10, random_state=0))  
    # Replaces missing values by fitting a Random Forest model

We’ll apply these techniques one by one to our dataset and see how the distribution of the data changes after the imputation. The code for this process is given in the section above.

We’ll also print the number of missing values after each imputation using the same isnull().sum() function we used earlier.

print(f"After {name} Imputation, Missing Values:\n{df_imputed.isnull().sum()}\n")

Remember, the goal of this exercise is not to eliminate missing values at any cost, but rather to handle them in such a way that we can still extract useful information from the data. That’s why we’re comparing the original distribution with the distributions after each imputation. We’re looking for an imputation technique that maintains the original data distribution as much as possible.

Visualizing the Treated Data

Last but not least, we visualize the treated data using histograms again. This helps us to compare the effects of different imputation techniques visually.

Remember, our goal here is not to achieve a perfect distribution after imputation, but rather to find a balance that minimizes the distortion while addressing the missing values.

By the end of this practical implementation, you should have a good understanding of how to handle missing values in real-world datasets. Remember, the best way to improve your skills is to practice. So don’t be afraid to get your hands dirty with other datasets and try out different imputation techniques!


Since Trinket Env does not support KNN Imputer and Regression Imputation, you will not be able to run this code in the IDE. You can find the code with KNN and Regression Imputation below.
Complete Code (Run this on your local Env or Google Colab Env)
from sklearn.experimental import enable_iterative_imputer
from sklearn.impute import KNNImputer, IterativeImputer
from sklearn.ensemble import ExtraTreesRegressor

# Define a dictionary of missing value handling techniques
techniques = {
    'Mean': SimpleImputer(strategy='mean'), 
    'Median': SimpleImputer(strategy='median'),
    'Most Frequent': SimpleImputer(strategy='most_frequent'),
    'KNN': KNNImputer(n_neighbors=3),  # Replaces missing values using the mean of 3 nearest neighbors
    'Iterative': IterativeImputer(estimator=ExtraTreesRegressor(n_estimators=10, random_state=0))  
    # Replaces missing values by fitting a Random Forest model

# We'll create a 3x2 grid of subplots, with one subplot for each imputation technique
fig, axs = plt.subplots(3, 2, figsize=(24, 21))
axs = axs.flatten()

# Plot the distribution of the 'sepal length (cm)' column before imputation
sns.histplot(df['sepal length (cm)'].dropna(), kde=True, ax=axs[0], color='blue', binwidth=0.2)
axs[0].set_title('Original Distribution', fontsize=16)

# Now we apply each imputation technique and visualize the results
for i, (name, imputer) in enumerate(techniques.items(), start=1):
    # Apply the imputer to the DataFrame
    df_imputed = pd.DataFrame(imputer.fit_transform(df))
    df_imputed.columns = df.columns
    df_imputed.index = df.index

    # Print the number of missing values after imputation
    print(f"After {name} Imputation, Missing Values:\n{df_imputed.isnull().sum()}\n")
    # Visualize the distribution of the 'sepal length (cm)' column after imputation
    sns.histplot(df_imputed['sepal length (cm)'], kde=True, ax=axs[i], color='blue', binwidth=0.2)
    axs[i].set_title(f'{name} Imputation', fontsize=16)

# Remove extra subplots
for i in range(len(techniques) + 1, 6):

# Make sure the subplots do not overlap


VIII. Improving Missing Values Handling: Considerations and Techniques

Handling missing values is an essential part of data preprocessing, but just as we don’t use a single model for all problems, there isn’t a one-size-fits-all solution for missing values. The handling technique depends on the type, distribution, and amount of data, as well as the specific scenario. Let’s take a look at how we can improve missing values handling:

1. Detecting the Nature of Missingness

Understanding the nature of the missingness in your data is crucial before deciding on a treatment method. There are three types of missingness: Missing Completely At Random (MCAR), Missing At Random (MAR), and Missing Not At Random (MNAR). Recognizing these types is not always easy, but getting this right can significantly improve your handling of missing data.

As a rule of thumb:

  • If the missingness is MCAR, listwise deletion can be used without introducing bias. However, this method should be used sparingly as it can waste valuable data.
  • If the missingness is MAR, it means there is a relationship between the propensity of a value to be missing and the observed data. In this case, techniques such as multiple imputation can be effective.
  • If the missingness is MNAR, the fact that a value is missing is related to the unobserved data (the missing value itself). This is the most challenging case, and you may need more advanced methods, like modeling the missingness or collecting more data.

2. Choosing the Right Imputation Technique

Imputation is a method to fill in the missing values with estimated ones. The objective is to employ known relationships that can be identified in the valid values of the data set to assist in estimating the missing values. Remember, the imputation method you choose should align with the nature of your data and the specific problem you’re trying to solve.

For instance, if you’re dealing with time series data, it might make sense to use methods like forward fill or backward fill, which fills the missing values with the preceding or succeeding values respectively. If the missing data is numerical, using mean, median, or mode to replace the missing values might be suitable. For categorical data, most frequent or a constant fill could be useful. More advanced techniques like K-Nearest Neighbors (KNN) and Multiple Imputations by Chained Equations (MICE) can also be considered for more complex scenarios.

3. Dealing with Categorical Variables

Categorical variables pose a unique challenge when it comes to handling missing values. For instance, replacing missing values with the most common category might introduce bias if the number of missing values is significant.

One approach to handling missing categorical values is to treat missing values as a separate category. This can be especially meaningful if the fact that the value is missing carries some information.

Another method is using algorithms like KNN and Random Forest that can handle missing values naturally. Alternatively, you could also use predictive modeling to fill missing categories, where a model is created to predict the missing values.

4. Feature Scaling: Addressing Skewed Data

Finally, addressing missing values might affect the distribution of your data. This is especially true if you have used a technique like mean or median imputation. After handling missing values, it’s important to revisit the distribution of your data. If the data is skewed, you might need to apply transformations or scaling methods to bring it to a normal or near-normal distribution. Techniques such as log transformation, square root transformation, or Box-Cox transformation can be used to remove skewness.

Also, feature scaling can be essential after treating missing values, especially if you’re using machine learning algorithms that are sensitive to the scale of the features, like Linear Regression, Logistic Regression, Support Vector Machines (SVM), KNN, K-Means, etc.

In conclusion, while handling missing values can seem daunting at first, a thorough understanding of your data and the problem context, combined with a knowledge of the various techniques available, can help you navigate this challenge effectively. Remember, the goal is not just to get rid of missing values, but to do so in a way that results in the most accurate and reliable machine learning model.

Remember, there’s no one-size-fits-all solution, and you might need to try different techniques and see what works best for your specific case. The key is to experiment and learn as you go!

IX. Applications of Missing Values Handling in Real World

Missing data is a common occurrence in real-world datasets. They can occur in a variety of fields, from healthcare and finance to e-commerce and social sciences. In this section, we will explore several real-world examples of missing values handling, examine its effects on model performance, and provide some guidance on when to choose which missing values handling method.

Real-World Examples of Missing Values Handling

Let’s take a look at a few scenarios where missing values handling plays a critical role.


In medical research, missing data is a frequent problem. Suppose we are working on a project studying the impact of various factors on heart disease. Certain critical features such as cholesterol levels or blood pressure may contain missing values. These could be due to several reasons: perhaps the patient didn’t show up for the test, or the results were lost or not recorded correctly.

In such cases, simply removing these records might result in a loss of valuable information. Instead, techniques such as mean or median imputation could be used if the data is missing completely at random. For more complex scenarios, regression imputation or multiple imputation might be better options.


Imagine we’re building a model to predict stock prices based on various economic indicators. There may be cases where some indicators are missing for certain periods. Deleting these records could result in inaccurate models, as it might ignore the time series nature of the data.

In such situations, it might be more appropriate to use methods like forward fill or backward fill, which fill the missing values with the preceding or succeeding values respectively. This maintains the continuity of the data, which is crucial for time series analysis.


In e-commerce datasets, customer reviews or ratings could be missing for some products. Ignoring these records might skew our analysis towards products with more reviews. However, treating these missing values as a separate category can be informative, as the fact that a product doesn’t have reviews could itself be an indicator of its popularity or lack thereof.

Effect of Handling Missing Values on Model Performance

The handling of missing values can significantly impact the performance of your machine learning models. Ignoring missing values may lead to biased or inaccurate models. On the other hand, different techniques for handling missing values can lead to varying model performance.

In general, more advanced techniques like KNN imputation and multiple imputation tend to result in more accurate models than basic techniques like mean or median imputation. However, this is not always the case and can depend on the nature and amount of missingness in your data. Therefore, it is recommended to try different methods and choose the one that results in the best model performance for your specific case.

When to Choose Which Missing Values Handling Method: Use Case Scenarios

Choosing the right missing values handling method depends on a variety of factors:

  1. Type of missingness: If the data is missing completely at random, techniques like listwise deletion or mean imputation could be used without introducing bias. However, for missing at random or missing not at random, more sophisticated techniques like multiple imputation or KNN imputation might be more appropriate.
  2. Type of data: If the missing data is numerical, using mean, median, or mode to replace the missing values might be suitable. For categorical data, creating a new category for missing values or using most frequent or a constant fill could be useful.
  3. Amount of missingness: If only a small percentage of values are missing, listwise deletion might be a reasonable approach. However, if a significant amount of data is missing, this could result in a loss of valuable information, and imputation methods might be a better option.

In conclusion, missing values handling is a critical aspect of data preprocessing in machine learning. It is a common issue in real-world datasets, and how we handle it can significantly affect our analysis and model performance. The key is to understand the nature and context of your data, and choose the most appropriate method for your specific situation.

X. Cautions and Best Practices with Missing Values Handling

Working with datasets that contain missing values can be tricky. While it’s essential to make strategic choices about how to handle these absences, it’s equally important to be aware of the potential pitfalls and follow best practices. Let’s delve deeper into the guidelines for handling missing data.

When to Use Which Missing Values Handling Method

The appropriate method for handling missing values largely depends on the nature of the missingness and the type of data.

  • Missing Completely at Random (MCAR): If the data is MCAR, meaning the missingness has no relationship with any values, observed or missing, simple methods such as listwise or pairwise deletion or mean, median, or mode imputation can be safely applied without introducing bias.
  • Missing at Random (MAR): For MAR data, where the propensity for a data point to be missing is not related to the missing data, but it is related to some of the observed data, more advanced techniques such as regression imputation, KNN imputation, or multiple imputation can be useful.
  • Missing Not at Random (MNAR): If the missingness is related to the missing data itself, careful handling is required. Advanced techniques like multiple imputation can be beneficial.

Also, take the type of data into consideration. For numerical data, mean, median, or KNN imputation might work well. For categorical data, consider creating a new category for missing values, or using most frequent or a constant fill.

When Not to Use Certain Missing Values Handling Methods

Not all missing value handling techniques are suitable for every situation. For instance:

  • Listwise or Pairwise Deletion: While simple to implement, these techniques should not be used if the percentage of missing data is high, as it might result in a significant loss of information.
  • Mean, Median, Mode Imputation: These methods can introduce bias or distort the distribution of the data, especially if the data is not MCAR or if the missingness is substantial.
  • Regression Imputation: While it can help to maintain relationships between variables, it might lead to an underestimation of errors.

Managing Impact on Variance and Distribution

Improper handling of missing values can distort the variance and distribution of the data, leading to biased or inaccurate models.

  • When using mean, median, or mode imputation, be aware that these methods might reduce variance and distort the distribution of the data, especially if the proportion of missing data is high.
  • Regression imputation maintains the variance but might underestimate the errors.
  • Advanced techniques such as multiple imputation can provide a better estimate of the variance and distribution, but they are more complex and computationally intensive.

Implications of Missing Values Handling on Machine Learning Models

The handling of missing values has direct implications on the performance of machine learning models.

  • Models trained on data with improperly handled missing values can result in biased or less accurate predictions.
  • Different machine learning algorithms have different sensitivities to missing data. For instance, algorithms like Decision Trees and Random Forests can handle missing values without any additional preprocessing, while algorithms like SVMs, Neural Networks, and Logistic Regression require the missing values to be handled.

Tips for Effective Data Preprocessing for Missing Values Handling

Here are some general best practices to follow when handling missing data:

  1. Perform a thorough exploratory data analysis (EDA): Understand the nature of your data, identify the patterns of missingness, and examine the relationships between variables.
  2. Choose the right method for your data and problem: Different methods have different assumptions and implications. Choose a method that is appropriate for your type of missingness and the nature of your data.
  3. Experiment with different methods: There’s no one-size-fits-all solution. Try out different methods and compare the results.
  4. Ensure reproducibility: Keep track of your preprocessing steps and make sure your process is reproducible.
  5. Document your findings: Keep a record of your findings from your EDA and the steps taken for handling missing data.

By following these guidelines and best practices, you can handle missing values effectively, minimize bias, and build robust machine-learning models.

XI. Missing Values Handling with Advanced Machine Learning Models

Handling missing values is a vital step in any machine-learning project. While basic and advanced imputation techniques can fill the gaps left by missing data, some advanced machine learning models have built-in mechanisms to handle missing values. In this section, we’ll dive deep into how missing values are treated in various advanced machine learning models, specifically focusing on Regression models, Classification models, and Deep Learning models.

How Missing Values Handling is Used in Regression Models

Regression models are used to predict continuous values. Traditionally, regression models, such as Linear Regression or Logistic Regression, cannot handle missing values and require complete data. This means before using these models, we need to use an imputation technique to fill in missing values. However, some advanced regression methods can deal with missing data:

  1. Multiple Imputation by Chained Equations (MICE): MICE is a method that fills missing values multiple times, creating “complete” datasets. Each “completed” dataset is then analyzed with standard procedures as if it had no missing values, and the results from these analyses are combined to produce estimates and confidence intervals. MICE is ideal for datasets where the missingness is random (MAR).
  2. Bayesian Linear Regression: Bayesian methods approach the problem of missing data differently by modeling the data generation process and integrating over the missing values. Bayesian Linear Regression can handle missing data by integrating the missing values out of the likelihood, which is done automatically when using Markov chain Monte Carlo (MCMC) methods.

Incorporating Missing Values Handling into Classification Models

Classification models predict categorical class labels. Some advanced classification models can handle missing values:

  1. Decision Trees and Random Forests: These models handle missing values effectively during both training and prediction. For example, when data is missing in a branching attribute, Random Forest will try both branches and weigh the results based on the proportions in the training data.
  2. XGBoost: XGBoost, a gradient boosting model, has a built-in routine to handle missing values. Upon reaching a node with missing values in the branching attribute, XGBoost algorithm takes the missing values and assigns them to whichever direction that optimizes the loss function.

The Interaction between Missing Values Handling and Deep Learning Models

Deep Learning models, like neural networks, typically require numeric inputs and cannot handle missing values. However, recent advances have allowed certain methods of handling missing data:

  1. DataWig: This is a library that learns Machine Learning models using Deep Learning to impute missing values in a DataFrame. It’s suitable for datasets where rows have some structure that can be captured by a deep-learning model.
  2. MissingNet: MissingNet is a deep learning methodology to handle missing values via denoising autoencoders. It imputes missing data by learning the structure of the data.

As we can see, different machine-learning models have different ways of dealing with missing values. In some cases, the model may handle missing values internally. In others, we will need to perform imputation before feeding data into the model. The choice depends on the specifics of our dataset and the problem we are trying to solve.

Remember, whatever method we choose, we need to consider the impact of missing values and our chosen handling method on model performance. Always validate your model using a hold-out test set to ensure that your approach to handling missing values is improving model performance.

XII. Summary and Conclusion

Recap of Key Points

In this comprehensive article, we explored the fundamental concepts, mathematical foundations, and practical applications of handling missing values in datasets – a critical aspect of data preprocessing in machine learning.

What are Missing Values? In the simplest terms, missing values are the data points that are absent in the dataset. These could be a result of various factors such as data entry errors, omissions, or intentional skipping of input fields by the data source.

Types of Missing Values: Missing data is typically classified into three types: Missing Completely At Random (MCAR), Missing At Random (MAR), and Missing Not At Random (MNAR). Understanding the nature of missingness helps guide the suitable method to handle the absence.

Implications of Ignoring Missing Values: Ignoring missing values can lead to significant consequences such as distorted statistical results, inaccurate machine learning models, and altered data distribution and variance.

Handling Missing Values: We discussed both basic and advanced techniques for handling missing values. Basic techniques included listwise and pairwise deletion, mean, median, and mode imputation, and random sampling imputation. Advanced techniques ranged from regression imputation and K-Nearest Neighbors (KNN) imputation to multiple imputation.

Practical Implementation: We walked through the practical implementation of handling missing values in Python, starting from data exploration and visualization to the final step of visualizing the treated data.

Improvements and Considerations: The article shed light on how to improve missing values handling, including detecting the nature of missingness, choosing the right imputation technique, dealing with categorical variables, and addressing skewed data through feature scaling.

Real-World Applications and Cautions: We illustrated real-world applications across various industries and cautionary advice on when to use (and not to use) certain missing values handling methods.

Missing Values Handling with Advanced ML Models: Finally, we discussed how missing values handling is incorporated in advanced machine learning models such as regression models, classification models, and deep learning models.

Closing Thoughts on the Importance of Handling Missing Values in Machine Learning

Addressing missing values is an indispensable step in preparing data for machine learning. The effectiveness of your chosen missing values handling method can significantly impact the performance of machine learning models. As such, it’s important to understand the nature and type of the missing values in your dataset and choose the most appropriate method for dealing with them.

Future Trends and Developments in Missing Values Handling Techniques

As machine learning continues to advance, we can expect more sophisticated methods for handling missing values to emerge. Techniques like the use of deep learning models (DataWig, MissingNet) to impute missing values show promise. The future may bring more refined techniques, perhaps ones that can handle missing values in a way that captures even more complex data patterns.

Through this journey, we have seen how handling missing values is much more than just ‘filling in the blanks’. It’s a careful process that requires consideration of the data, the missingness, and the potential impact on our models. It’s part science, part art, and entirely crucial in the realm of machine learning. Whether you’re a novice data scientist or a seasoned machine learning engineer, understanding and effectively handling missing values will invariably be an essential part of your toolkit.

Share the Post:
Learn Data Science. Courses starting at $12.99.

Related Posts

© Let’s Data Science


Unlock AI & Data Science treasures. Log in!