TakeAPeek Learning: Chaining classifiers for multiple regression

Abstract

In multiple target regression problems, the objective is to predict the numeric value of multiple dependent Y variables from the numeric values of independent X variables using linear regression. One of the issues with this type of problem is the fact that it does not account for the codependency between the target variables Y themselves. This objective of this thesis is to adapt the existent multiple regression algorithm to use chained classifiers. By using chained classifiers the prediction of each target variable Yn is not done independently of the other Y variables. Each prediction model will include the values of previously predicted Y variables. By doing so the model will attempt to combat the co dependence between the target variables thus leading to more accurate results. This thesis works upon an existing article about multi-label classification methods for multi-target regression. I will demonstrate the effectiveness of the base algorithm using 12 datasets and prove the influence that the order in which the target variables are predicted have upon the final mean error of all target variables.