What are the best practices for handling missing data in supervised learning models?

Missing data is a common and inevitable problem in many supervised learning models, especially when dealing with real-world datasets. It can affect the quality, accuracy, and performance of the models, and introduce bias and uncertainty in the results. Therefore, it is important to handle missing data properly and follow some best practices to avoid or minimize its negative impact. In this article, we will discuss some of the best practices for handling missing data in supervised learning models, such as identifying the types and patterns of missing data, choosing appropriate methods to deal with missing data, and evaluating the effects of missing data on the models.

1 Identify missing data

The first step to handle missing data is to identify its presence, extent, and nature in the dataset. You can use various tools and techniques to explore and visualize the missing data, such as summary statistics, heatmaps, histograms, and scatter plots. You should also try to understand the reasons and mechanisms behind the missing data, as they can influence the choice of the methods to deal with it. There are three main types of missing data mechanisms: missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR). MCAR means that the probability of missing data is independent of any variable in the dataset. MAR means that the probability of missing data depends on some observed variables, but not on the missing values themselves. MNAR means that the probability of missing data depends on the missing values or some unobserved variables.

Add your perspective

Rômulo B.

Head of Information Technology | CTO | Leading Technology Strategy | Data Strategy
Report contribution
We could define several ways to help identify missing data, such as: - Exploration data analysis: Begin by visually exploring the data, using histograms and statistics to identify missing data. Libraries like Matplotlib, Pandas, and Seaborn are good options in Python to help in this exploration. - Missing values count is a good strategy as well, because could you use functions like isnull() or isna() to check the number of missing values in the data. - One of the strategies would be to use a couple of algorithms, such as Random Forest e XGBoost that can help you to identify automatically detect missing values during training.

Like
Karthikheyaa Kurra

Machine Learning Practitioner | Grad Student at UD
Report contribution
One thing I've found helpful is researching your dataset (struct or unstruct). This is one of the underrated qualities that is overlooked by many beginners (I did too). One time, I was dealing with a horse's health dataset in a Kaggle competition and I was kind of struggling to fill in the missing data I just filled it with mean/median/mode and got the worst model score. So I started researching more about what the columns mean and how they are dependent on each other in the real world, and I realized that Nan in one col can be filled based on other cols (for ex: I filled the heart condition col based on the obesity col and so on). So the point is, try to use a combination of Domain Knowledge and other techniques to boost your model score.

Like
Osama Ammar, MSc.

Computer Vision Engineer @Atomica.ai | AI | ML | MLOps | Medical Imaging
Report contribution
Statistical analysis of your dataset plus visualization are key steps before doing anything such as training or model selection.

Like

2 Choose appropriate methods

The next step is to choose appropriate methods to deal with the missing data, depending on the type, pattern, and amount of missing data, as well as the goal and complexity of the supervised learning model. There are two main categories of methods: deletion and imputation. Deletion means removing the observations or variables that contain missing data from the dataset. This can be done by listwise deletion, which removes any row that has at least one missing value, or pairwise deletion, which removes only the pairs of values that are missing for a specific analysis. Deletion is simple and easy to implement, but it can reduce the sample size, introduce bias, and distort the relationships among variables. Imputation means replacing the missing values with some estimated values based on the available data. This can be done by various methods, such as mean, median, mode, or constant imputation, which use a single value for all missing values, or regression, interpolation, or k-nearest neighbors imputation, which use more complex algorithms to estimate the missing values based on their relationships with other variables. Imputation can preserve the sample size, reduce bias, and improve accuracy, but it can also introduce noise, variance, and uncertainty in the data.

Add your perspective

Osama Ammar, MSc.

Computer Vision Engineer @Atomica.ai | AI | ML | MLOps | Medical Imaging
Report contribution
simple techniques in short : - Removal of rows/columns - Imputation (mean, median, mode) - Predictive modeling - Interpolation

Like
Chander Parkash

Creating new Software Engineers 👨💻 | Java | Go (Golang) | Spring Boot | Spring Core & MVC | TypeScript | Nest.js | Angular | Flutter | PostgreSQL
Report contribution
When addressing missing data, selecting appropriate methods hinges on factors like the type, pattern, and volume of missing data, as well as the objective and complexity of the supervised learning model. Primarily, two categories of methods exist: deletion and imputation. Deletion involves removing observations or variables containing missing data from the dataset. Listwise deletion eliminates any row with at least one missing value, while pairwise deletion targets only missing value pairs for specific analyses. While straightforward, deletion may shrink sample size, introduce bias, and distort variable relationships.

Like

3 Evaluate the effects

The final step is to evaluate the effects of missing data and the chosen methods on the supervised learning model. You should compare the performance and accuracy of the model with and without missing data, and with different methods of dealing with missing data. You should also assess the validity and reliability of the model results, and check for any potential bias or inconsistency caused by missing data. You can use various metrics and techniques to evaluate the effects of missing data, such as cross-validation, error analysis, sensitivity analysis, and confidence intervals. You should also report and document the missing data and the methods used to handle it, and discuss the limitations and implications of the missing data for the model and the domain.

Add your perspective

Deepak Kumar

3x Founder | Building Vyaan.io | AI SaaS | GTM | B2B SaaS | IIT Patna
Report contribution
Navigating missing data in supervised learning is a three-step dance: start by spotting it through histograms and stats, know its type (MCAR, MAR, MNAR). Next, choose your move—delete or impute (mean, median, k-nearest neighbors). Finally, hit the floor, evaluate effects with cross-validation, and report your steps. It's a rhythmic journey in the world of supervised learning.

Like

4 Here’s what else to consider

This is a space to share examples, stories, or insights that don’t fit into any of the previous sections. What else would you like to add?

Add your perspective

Deepak Kumar

3x Founder | Building Vyaan.io | AI SaaS | GTM | B2B SaaS | IIT Patna
Report contribution
Beyond handling missing data, consider it as a dance of data quality. Dive into the reasons behind gaps, whether it's tech glitches or the quirks of collection. Team up with data wizards, domain experts, and collectors for a symphony of insights. Keep the conversation flowing to refine strategies and create models that groove with the dynamic nuances of real-world data. It's not just about handling gaps; it's about crafting a resilient data narrative.

Like

What are the best practices for handling missing data in supervised learning models?

1

2

3

4

1 Identify missing data

2 Choose appropriate methods

3 Evaluate the effects

4 Here’s what else to consider

Computer Science

Rate this article

Thanks for your feedback

More articles on Computer Science

More relevant reading

What are the best practices for handling missing data in supervised learning models?

1

2

3

4

1 Identify missing data

2 Choose appropriate methods

3 Evaluate the effects

4 Here’s what else to consider

Computer Science

Rate this article

Thanks for your feedback

Explore Other Skills