Multiclass, multilabel, and multitask classification

Are you wondering what the difference between multiclass and multilabel classification is? Or maybe you are interested in hearing about the difference between multilabel and multitask classification? Well either way, you are in the right place! In this article we tell you everything you need to know to understand what multiclass, multilabel, and multitask classification are and how they differ from one another.

We start out by reviewing multiclass, multilabel, and multitask classification frameworks one at a time to establish a clear understanding of what each framework is. This overview includes real world examples of classification problems that are appropriate for each framework. After that, we discuss the similarities and differences between these classification frameworks.

Outcomes, levels, and labels

Before we dive into our discussion, we will first take a step back and define a few different terms we will be using throughout this article.

• Outcome variable. An outcome is a distinct variable or type of attribute that you want to predict. A given outcome variable can take on any number of values, but all of the values should be logically grouped or in some way related to one other. Throughout this article we will be discussing both models with a single outcome variable and models with multiple different outcome variables.
• Level. A level is just a value that a specific outcome variable can take on. For example, if your outcome variable is the color of an apple then a few examples of levels that outcome variable can take on are red, green, and multicolored.
• Label. Label is a word we will use to refer to a level of an outcome variable that has been applied to a given observation. In the simple situation where there is a single outcome variable from which a single level is chosen, there will always be one label per observation. However, in more complex cases such as cases where there are multiple different outcome variables or cases where multiple levels of a single outcome variable can be applied to a single observation, there may be many labels per observation.

What are multiclass, multilabel, and multitask classification?

What is multiclass classification?

We will start out by talking about multiclass classification. This is a great place to start because multiclass classification is the most common framework that comes up once you move past simple binary classification. Multiclass classification is actually very similar to binary classification in that there is a single outcome variable that can take on only one level at a time. The main difference between binary classification and multiclass classification is that in multiclass classification there are more than two possible levels that your outcome variable can take on.

Let’s introduce a simple example to help you understand the types of problems multiclass classification can help solve. Imagine that you wanted to develop a model that used information such as color and weight of an apple to predict the type of apple. In this situation your outcome variable would be the type of apple and the levels the outcome variable can take on might be something like gala, fuji, and granny smith. An apple has one and only one type, so exactly one label should be applied to each observation in your dataset. This is a perfect example of multiclass classification.

What is multilabel classification?

The next framework that we will talk about is multilabel classification. This is another framework that can be used when you have one outcome variable that has multiple levels. So what differentiates multilabel classification from other types of classification? Multilabel classification is used when you want to allow more flexibility in the number of labels that can be applied to a given observation rather than limiting yourself to strictly one label per observation. With multipliable classification you can select zero, one, or multiple different labels that apply to a given observation.

Let’s modify the example that we used in the previous section so that it is more appropriate for multilabel classification. We will stick to the apple theme, but this time we will say that we want to identify what defects an apple has. Some examples of levels this outcome variable might take on are overripe, bruised, and undersized. This case is different from the previous case because a given apple can have all of these attributes, none of these attributes, or any given subset of the attributes. This is a perfect example of a multilabel classification problem.

What is multitask classification?

So what is multitask classification? Multitask classification a framework you can use if you have multiple different outcome variables you want to consider. Specifically, you should use multiclass classification if you have multiple different outcome variables you want to predict using the same model. This is generally appropriate if your outcome variables are interdependent and knowing information about one of the variables provides information about the other.

Let’s again modify our apple example so that it is appropriate for a multitask classification model. Imagine now that you are given a picture of an apple and you want to get more information about the size and weight of the apple. In this case, your first outcome is size and your second outcome is weight. This is a perfect example of a multitask classification problem because there are two different outcome variables that are conceivably related to one another.

Differences between classification methods

What is the difference between multiclass and multilabel classification?

What is the difference between multiclass and multilabel classification? Multiclass and multilabel classification are both used in situations where you have a single outcome variable that has multiple different levels. The main difference between multiclass and multilabel classification is the number of labels that are applied to each observation. In multiclass classification, one label is applied to each observation. In multilabel classification, the list of labels applied to each observation can range from no labels to every possible label.

What is the difference between multiclass and multitask classification?

What is the difference between multiclass and multitask classification? Multiclass and multitask classification methods are both used in cases where you want to apply one label per outcome variable. The main difference between multiclass and multitask classification is the number of outcome variables in the model. In multiclass classification there is only one outcome variable, whereas in multitask classification there are multiple outcome variables that should be considered jointly.

What is the difference between multilabel and multitask classification?

What is the difference between multilabel and multitask classification? Multilabel and multitask classification are both methods that result in multiple labels being applied to each observation. The difference is how these labels are allocated. In multitask classification, there are multiple different outcome variables and one label is applied for each outcome variable. In multilabel classification, there is one outcome variable but multiple labels can be applied for that outcome variable.