Imputación de datos incompletos y clasificación de patrones mediante aprendizaje multitarea

Abstract

Almost all research on supervised learning is based on the assumption that training data are completely observable, but it is not a common situation because real world databases are rarely complete. The ability of handling missing data has become a fundamental requirement for machine learning. Up to now, proposed methods consider the problem as two separated tasks, main task and imputation task, and solve them separately (Single Task Learning, STL). In this paper, a new effective method is proposed to handle missing features in incomplete databases with Multitask Learning (MTL). This approach uses the imputation task as extra task and learning in parallel with the main task. Thus, imputation is guided and oriented by the learning process, i.e., imputed values are those that contribute to improve the learning. In this paper we use the advantages of MTL to handling missing data and analyze its robustness for handling different missing variables in real an artificial data sets.