| dc.description.abstract | This thesis conducts a rigorous examination of the potential for Variational Autoencoder (VAE)-derived latent embeddings to enhance the performance of classification algorithms when assessing business risk profiles from accounting data, specifically focusing on the Norwegian context. The study compares the performance of classifiers using VAE latent embeddings against those utilizing original or balanced training sets directly. Generally, it was found that, with the exception of Logistic Regression in certain experimental settings, the performance of classifiers using VAE latent embeddings was somewhat inferior. This outcome suggests that the dimensionality reduction process inherent to VAE may induce a degree of predictive power loss.
However, VAE latent embeddings were observed to bolster the performance of Logistic Regression by effectively capturing complex, high-dimensional relationships within a compressed, lower-dimensional space. This process reduced noise, identified non-linear relationships, and introduced a beneficial regularization effect, which may enhance the generalizability of the Logistic Regression model.
Furthermore, an increase in the dimensionality of the latent space up to a certain threshold improved the performance of classifiers, beyond which a decline was observed, indicating an optimal dimensionality for these datasets. The application of under-sampling or over-sampling techniques to the training sets generally led to decreased classifier performance, particularly for Extreme Gradient Boosting and Multi-Layer Perceptron, with Logistic Regression as an exception in certain contexts.
Notably, for the Norwegian dataset, the Extreme Gradient Boosting classifier often demonstrated superior performance when utilizing raw training sets. These findings provide valuable insights into the capabilities and limitations of VAE in assessing business risk profiles and underscore the need for further research in this promising field. | en_US |
