Using VAE for Incomplete Educational Data

• Ano de defesa: 2023
• Autor(a) principal: Montecino, Claudia Evelyn Escobar
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Autoencoder; Autoencoder variacional; Dados ausentes; Dados educacionais incompletos; Incomplete educational data; Item response theory; Missing data; Neural networks; Redes neurais; Teoria da resposta ao item; Variational autoencoder
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/104/104131/tde-24082023-102049/

In Psychometrics, and in particular in educational assessments, it is common to find incomplete databases. Lack of time, forgetting the content involved, nervousness or even the test design are some of the reasons why an individual may leave items unanswered in an assessment. In this context, it is important to have estimation methods for psychometric models that deal with missing data and are affected as little as possible by the lack of information in those unanswered items. In a small-scale scenario, traditional estimation methods for Item Response Theory (IRT) models, for example, are suitable for situations with complete and incomplete data. However, for high-dimensional situations, such as assessments involving many latent skills and abilities, traditional methods are not computationally efficient or even unable to obtain estimates for so many parameters. Deep learning has been adapted to incorporate IRT models and make predictions and estimates from large, high-dimensional databases. In this work, we deepen the investigation of (?)]Curi, who defined a Two Parameter Logistic Model (ML2P) in the architecture of a variational autoencoder (VAE) as a proposal to solve the problem of estimating the many parameters of the model. We performed a simulation study to compare two variations of deep neural networks, autoencoders (AE) and VAE, defined with an ML2P model in the decoder, for situations with a large number of latent traces and complete data. After favorable results of the VAE, we propose an extension of the same (IVAE) to be able to make predictions in cases of missing data and, thus, make the model more general and useful in practice. Simulations of the proposed model were performed under different scenarios to investigate the efficiency of the new method in recovering the parameters. Comparisons of the results with one of the methodologies currently most indicated in IRT to deal with a situation of greater dimensionality, the joint maximum likelihood, were also made, in addition to the application to a real case of high dimension and with missing data.