Author: Jingjun Bi

Thesis: "Contribution to Data Augmentation for Image Classification and Segmentation"

Director: Fadi Dornaika

Day: 2024ko uztailaren 30ean
Time: 10:30h
Place: Ada Lovelace aretoa

Abstract:

"In recent years, semi-supervised learning on graphs has garnered significant attention across various domains and applications. The objective is to leverage both partially labeled data (consisting of labeled examples) and a vast pool of unlabeled data to develop more robust predictive models. Deep Graph Neural Networks (GNNs) have emerged as powerful tools for addressing both unsupervised and semi-supervised learning tasks. Among these, Graph Convolutional Networks (GCNs), as a specialized class of GNNs, aim to extract meaningful data representations through graph-based node smoothing and layer-wise neural network transformations.

While certain scenarios inherently exhibit graph structures in their data, such as social networks or interconnected objects, many other types of data lack such inherent graph structures. For instance, image data often entails diverse features describing each image, representing a typical form of multi-view data. Despite the versatility of GCNs, there has been a notable gap in deep learning approaches specifically tailored for multi-view graph-based semi-supervised learning, particularly for datasets lacking inherent graph structures.

In this thesis, we focus on enhancing the efficacy of Graph Convolutional Neural Networks on both single-view and multi-view datasets, focusing particularly on datasets devoid of explicit graph structures. We scrutinize the existing limitations of GCNs when applied to single-view semi-supervised graph learning tasks. To address these shortcomings, we introduce a novel semi-supervised learning method termed Re-weight Nodes and Graph Learning Convolutional Network with Manifold Regularization (ReNode-GLCNMR). This method amalgamates graph learning and graph convolution within a unified network architecture, while also incorporating label smoothing through an unsupervised loss term. Moreover, it tackles the challenge of imbalanced graph topology by adaptively reweighting the influence of labeled nodes based on their proximity to class boundaries.

However, there exists a notable gap in applying Graph Convolutional Networks (GCNs) to multi-view data lacking explicit graph structures. To bridge this gap, we introduce four novel deep semi-supervised multi-view classification models explicitly designed for non-graph data.

The first model, termed Semi-supervised Classification with a Unified Graph (SCUG), and the second model, Semi-supervised Classification with a Fused Graph (SCFG), share a common architecture leveraging GCNs and incorporating a label smoothing constraint. The primary distinction lies in the construction of the consensus similarity graph. In SCUG, the consensus graph is directly reconstructed from the different views using a specialized objective function tailored for flexible graph-based semi-supervised classification. Conversely, SCFG independently reconstructs individual graphs before adaptively merging them into a unified consensus graph. The third model, Sample-weighted Fused Graph-based Semi-supervised Classification (WFGSC), is devised for multi-view data and follows several key steps: (i) constructing a semi-supervised graph in each view through a flexible model jointly estimating graphs and labels, (ii) generating an additional graph based on node representations provided by the joint estimator, (iii) assigning higher weights to hard-to-classify samples, and (iv) proposing a loss function integrating the graph auto-encoder loss and label smoothing over the consensus graph for training the GCN on fused features.

To enhance efficiency, we refine WFGSC by utilizing a single GCN to achieve all features, eliminating the need for two separate GCNs. Finally, we introduce the fourth model, Linear Projection Fused Graph-based Semi-supervised Classification (LFGSC). This approach initially employs a semi-supervised method for each view to concurrently estimate corresponding graphs and flexible linear data representations in a low-dimensional feature space. Subsequently, an adaptive and unified graph is generated, followed by leveraging a fully connected network to fuse the projected features and reduce dimensionality. Finally, the fused features and graph are inputted into a GCN for semi-supervised classification. During training, the model accounts for cross-entropy loss, manifold regularization loss, graph auto-encoder loss, and supervised contrastive loss. By employing linear transformation, input feature dimensions for the GCN are significantly reduced, achieving high accuracy while mitigating computational costs.

Moreover, extensive experimental results conducted on various benchmark datasets underscore the superiority of our proposed methods.

Keywords: Semi-supervised learning, Deep Graph Neural Networks, Graph Convolutional Networks, Graph construction, Graph fusion, Graph regularization, Multi-view data."