Year: 2025
Type: dissertation
Abstract: The continuous growth of machine learning methods has driven significant advances in areas such as natural language processing and computer vision. However, the performance of these models heavily depends on the quantity and quality of labeled data available for training, which is often scarce or expensive to produce, especially in specialized domains. This challenge becomes even more critical in the face of the phenomenon of domain shift, characterized by the difference between the training data distribution and the distribution of the data where the model will be applied. These limitations pose significant barriers to the application of state-of-the-art models in real-world scenarios. Domain adaptation emerges as a solution to mitigate the impacts of domain shift, optimizing the transfer of learning between different domains. This field aims to create domain-invariant representations that preserve relevant information for a specific task, thereby reducing the discrepancy between data distributions. Despite notable advances, the field still faces challenges such as balancing the specificity and generalization of the generated representations, as well as issues related to the high dimensionality of projections in latent spaces, which result in high computational costs and low interpretability. In this thesis, we propose new approaches to address the challenges of domain adaptation, focusing on balancing generalization and specificity, increasing interpretability and computational efficiency, and improving model usability. The contributions include advancements applied to tasks such as sentiment analysis, image classification, and question-answering systems. We present the Cluster Fusion Domain Adaptation (CFDA) method, which combines clustering techniques and representation fusion to generate multiple representations that ensure both internal domain alignment and generalization across domains. To tackle the issues of computational efficiency and interpretability, we developed the Pseudo-labeling Domain Adaptation (PDA) method, which employs projections in low-dimensional spaces and graph neural networks to model relationships between domains. Additionally, to enhance usability in aspect-based sentiment analysis, we introduced a novel question-answering method that enables more flexible interactions between users and applications. The experimental results of this thesis validate the proposed methods, demonstrating improvements in accuracy, computational efficiency, and generalization capacity across different datasets. Thus, this thesis makes significant contributions to the domain adaptation field, offering more versatile, interpretable, and accurate methods for existing challenges.
Author: Victor Akihito Kamada Tomita
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4413040860