Item response theory: autoencoders

• Ano de defesa: 2024
• Autor(a) principal: Tabak, Gabriel Couto
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Algoritmo evolutivo; Autoencoders; Distribuições não normais; Evolutionary algorithm; Item response theory; Neural architecture search; Neural network; Non-normal distributions; Redes neurais; Simulações; Simulations; Teoria de resposta ao item
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-17032025-163109/

Autoencoders, a method of unsupervised learning, aim to capture lower-dimensional latent spaces. They consist of two interconnected neural networks: an encoder that condenses information into a compact representation and a decoder that reconstructs the original data from these compressed features. They are versatile tools for discovering latent representations in various contexts. This dissertation introduces and explores the application of autoencoders within Item Response Theory (IRT), proposing specific models that correspond to the logistic two-parameter model and the graded response model. The inherent flexibility of neural networks is expected to provide distinct advantages in the estimation of IRT parameters. Our study focuses on their efficacy in parameter retrieval comparing normal and non-normal distributions. To maintain parameter interpretability, we have fixed the decoders architecture for each model, and to address the issue of model identifiability, we have proposed two distinct constraints within the decoder. Our initial inquiry examined whether certain neural network architectures for the encoder (specific configurations of neurons and layers) were particularly effective using an evolutionary neural architecture search. Although no single architecture emerged as universally superior, the imposed decoder constraints proved sufficient for consistent parameter estimation across various structures. Nonetheless, challenges appeared in scenarios characterized by some adversity. Items with low discrimination or skewed response distributions (excessive zeros or ones) impact the precision of latent trait retrieval. Comparative analysis of the mean bias in parameter retrieval revealed that while our autoencoder approach aligns with other methods to estimate IRT parameters, it does not surpass them in most metrics. Notably, in cases of non-normality, our method demonstrated robust estimation capabilities. One advantage of our proposed method was observed in the extremities of the latent trait distribution, where the autoencoder exhibited bias around zero. This contrasts with other methods that tended to exhibit positive bias in the lower tail and negative bias in the upper tail of the distribution.