Author:

Keywords:

machine learning, multi-target learning

Abstract:

In this work we will investigate multi-target Learning, also called multi-tasklearning, a branch of Machine Learning. A multi-target learner builds a modelwhich, given some input, returns/predicts multiple variables simultaneously.This stand in contrast with the usual Single-Target model which predicts onlyone variable. We will see that MT-models have numerous advantages and thatmany real world applications can benefit from using a multi-target approach. We observe that when multiple targets are predicted simultaneously by a MT-model, the predictive accuracy might increase in comparison to the single-targetmodel. This is a result of inductive transfer which we discuss in depth in chapter3. An algorithm is proposed which tries to maximally exploit this effect in orderto increase the predictive accuracy of the model for one specific target (themain target). Next, we focus our attention to a different problem: Collaborative Filtering, atype of recommender system. Such a system tries to predict user preferences, foritems such as books or movies, based on already known preferences of the users.We demonstrate how this problem fits the multi-target setting and proposea graph based algorithm to alleviate two common problems inherent to theCollaborative Filtering setting: the Cold Start and the Sparsity problem. Finally, we propose a new learning setting called Two-Way learning. A specifictype of multi-target learning where we try to predict an attribute of a relationbetween two types of objects. We propose data transposition as an effectivemethod to solve a Two-Way learning problem. This method is applied andstudied in depth on the problem of processor performance prediction.