What is the chained multioutput regressor?

In a multioutput regression problem, there is more than one target variable. These target variables are continuous variables. Some machine learning algorithms like linear regression, KNN regression, or Decision Tree regression can solve these multioutput regression problems inherently. But, some machine learning algorithms like the Support Vector Machine Regression (SVR) cannot solve multioutput regression problems inherently. We need to use a direct or chained multioutput regressor along with such machine learning algorithms to solve a multioutput regression problem.

In our previous article, we discussed direct multioutput regression. In chained multioutput regressor also a multioutput regression problem is broken into several subproblems. But, unlike direct multioutput regression, in chained multioutput regression the previously predicted target variables are used to predict the next target variable.

For example, let’s say we need to find out the x, y, and z values based on some features. A chained multioutput regressor breaks the multioutput regression problem into the following subproblems:

Problem 1: Given the features, predict x.
Problem 2: Given the features and the predicted x value, predict y.
Problem 3:  Given the features and the predicted x and y values, predict z.

Now, these subproblems can be solved using any regressor and all the target variables can be predicted.

Chained multioutput regression using sklearn in Python

We can use the following Python code to implement a chained multioutput regressor using linear SVR in sklearn. …