The paper presents a new discretization method applicable to measures of continuous latent traits estimated using a measurement model belonging to the item response theory (IRT) approach. The reasons of this proposal are twofold: first, the need to discretize a continuous variable due to the use of methodologies primarily designed to handle categorical data (for example Bayesian Networks) or the increase of efficiency and effectiveness of the learning algorithms, second, the discretizers available in literature are not able to reproduce the peculiarities of the target variables of this paper. The idea underlying the proposed method is to use the information from an IRT model in order to forecast the answer of a subject to a characterizing item; the obtained response is the category assigned to the subject in the discretized version of her/his continuous latent trait. The performance of this discretizer is compared to the performance of other common unsupervised discretization methods, with respect to a global single-item measure, that is assumed to represent an observed discretized version of the continuous latent trait.
A discretization method for continuous latent traits
Golia S.
2021-01-01
Abstract
The paper presents a new discretization method applicable to measures of continuous latent traits estimated using a measurement model belonging to the item response theory (IRT) approach. The reasons of this proposal are twofold: first, the need to discretize a continuous variable due to the use of methodologies primarily designed to handle categorical data (for example Bayesian Networks) or the increase of efficiency and effectiveness of the learning algorithms, second, the discretizers available in literature are not able to reproduce the peculiarities of the target variables of this paper. The idea underlying the proposed method is to use the information from an IRT model in order to forecast the answer of a subject to a characterizing item; the obtained response is the category assigned to the subject in the discretized version of her/his continuous latent trait. The performance of this discretizer is compared to the performance of other common unsupervised discretization methods, with respect to a global single-item measure, that is assumed to represent an observed discretized version of the continuous latent trait.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
