Leveraging Graph Convolutional Networks for Semi-Supervised Learning in Multi-View Non-Graph Data

Artificial Intelligence Computer Science Information Science
multi-view data semi-supervised learning graph estimation consensus graph graph convolutional networks

Background Introduction

In the field of machine learning, Semi-Supervised Learning (SSL) has garnered significant attention due to its ability to leverage a small amount of labeled data and a large amount of unlabeled data for learning. Particularly in scenarios where data labeling is costly, graph-based semi-supervised learning methods have become a research hotspot. Graph Convolutional Networks (GCNs) have shown remarkable performance in semi-supervised learning, especially in data with inherent graph structures, such as citation networks and social networks. However, there remains a noticeable gap in the application of GCNs to non-graph multi-view data, such as image collections.

Multi-view data refers to datasets that capture information about the same object from different perspectives or modalities. For example, TV data includes video and audio as two views, natural language understanding may express the same semantic object in different languages, and face recognition may represent facial data through 2D images and 3D models. Multi-view learning aims to utilize these complementary pieces of information to build a unified model, thereby enhancing classification performance. Nevertheless, existing multi-view learning methods still face challenges when dealing with non-graph data, particularly image data.

To address this issue, researchers F. Dornaika, J. Bi, and J. Charafeddine proposed a multi-view semi-supervised classification model based on GCNs, specifically tailored for non-graph data. Their research aims to fill this gap and provide new solutions for semi-supervised learning in multi-view data.

Source of the Paper

The paper was co-authored by F. Dornaika (University of the Basque Country and Ikerbasque Science Foundation), J. Bi (University of the Basque Country), and J. Charafeddine (De Vinci Higher Education Research Center). It was published in the 2025 issue of the journal Cognitive Computation, with the title Leveraging Graph Convolutional Networks for Semi-Supervised Learning in Multi-view Non-graph Data and the DOI 10.1007/s12559-025-10428-y.

Research Process

1. Research Objectives and Method Design

The primary objective of this research was to develop two multi-view semi-supervised classification models based on GCNs, named “Semi-Supervised Classification with a Unified Graph” (SCUG) and “Semi-Supervised Classification with a Fused Graph” (SC-Fused). Both models share the commonality of using the GCN framework and incorporating a Label Smoothing Constraint. Their distinction lies in the construction of the Consensus Similarity Graph.

2. Semi-Supervised Classification with a Unified Graph (SCUG)

The core idea of SCUG is to directly reconstruct the consensus graph from different views using a specialized objective function. The specific steps are as follows:

  1. Data Preprocessing: Normalize the features of samples in each view to ensure that the column vectors of the data matrix are unit vectors.
  2. Unified Graph Construction: Use the Multi-view Consistent Graph Construction and Label Propagation Algorithm (MVCGL) to estimate the unified graph. This algorithm optimizes a global objective function, leveraging labeled data and predicted labels to generate a discriminative semi-supervised model.
  3. GCN Training: Input the unified graph and global feature matrix into the GCN architecture. Train the model through layer-wise propagation and label smoothing constraints, ultimately outputting soft label predictions for all samples.

3. Semi-Supervised Classification with a Fused Graph (SC-Fused)

SC-Fused adopts an adaptive fusion approach to construct the unified graph. The specific steps are as follows:

  1. Individual Graph Construction: Build independent similarity graphs for each view by optimizing an objective function to generate the graph matrix for each view.
  2. Fused Graph Construction: Adaptively merge individual graphs into a unified consensus graph based on the data smoothness weights of each view.
  3. GCN Training: Input the fused graph and global feature matrix into the GCN architecture. Train the model through layer-wise propagation and label smoothing constraints, ultimately outputting soft label predictions for all samples.

4. Experimental Design and Datasets

To validate the effectiveness of the proposed models, the researchers conducted experiments on seven multi-view image datasets, including ORL, Scene, Handwritten, ALOI, MSRC-v1, YouTube, and MNIST. These datasets cover various types of image data, such as face images, scene images, and handwritten digit images.

5. Comparison Methods and Parameter Settings

The researchers compared SCUG and SC-Fused with seven existing methods, including two baseline methods (GCN-X* and GCN-Multi) and four state-of-the-art multi-view semi-supervised learning methods (MVCGL, AMSSL, DSRL, and JCD). The parameter settings for all models were kept consistent to ensure fairness in the experiments.

Research Results

The experimental results showed that SC-Fused achieved the highest classification accuracy on six datasets (ORL, Scene, Handwritten, ALOI, MSRC-v1, and YouTube), demonstrating significant superiority. SCUG performed exceptionally well on four datasets (Scene, ALOI, MSRC-v1, and YouTube), ranking second only to SC-Fused. In contrast, the performance of other methods varied significantly across different datasets, and they performed poorly on complex datasets.

1. Parameter Sensitivity Analysis

The researchers conducted a detailed analysis of the parameter sensitivity of SC-Fused, revealing significant differences in optimal parameter settings across datasets. For example, the optimal balance parameter λ for the ALOI dataset was 0.1, while for the Handwritten dataset, it was 1200. This indicates that different datasets require parameter adjustments based on their characteristics to achieve the best classification results.

2. Graph Construction and Classification Effectiveness

By visualizing the similarity matrices, the researchers found that SC-Fused effectively captured intra-class sample similarity and reduced inter-class sample similarity during graph matrix construction. This effective graph construction directly improved the accuracy of semi-supervised classification. For instance, on the Handwritten and ORL datasets, the graph matrices constructed by SC-Fused showed clear intra-class clustering and inter-class separation, aligning with their high classification accuracy.

3. Embedding Visualization

Using t-SNE visualization, the researchers demonstrated the distribution changes in the input features and output representations of the SC-Fused model. The results showed that semi-supervised learning made samples of the same class more clustered and samples of different classes more separated, further validating the model’s effectiveness.

Research Conclusions

The study proposed two multi-view semi-supervised classification models based on GCNs, filling the gap in the application of GCNs to non-graph data. The experimental results demonstrated that SC-Fused performed exceptionally well on multiple datasets, particularly showing significant advantages when handling complex datasets. The main contributions of the research include:

  1. Proposing two methods for constructing consensus graphs, applicable to multi-view and non-graph data.
  2. Using the generated graphs to train semi-supervised GCNs, enhancing classification performance.
  3. Validating the superiority of the proposed methods on multiple datasets through experiments.

Research Highlights

  1. Innovation: This study is the first to apply GCNs to non-graph multi-view data, proposing a novel semi-supervised classification framework.
  2. Effectiveness: SC-Fused achieved the highest classification accuracy on multiple datasets, demonstrating significant superiority.
  3. Application Value: The study provides new solutions for semi-supervised learning in multi-view data, with broad application prospects, particularly in image classification, video analysis, and natural language processing.

Future Prospects

The researchers indicated that future work will focus on reducing the computational complexity of multi-view data, especially when dealing with high-dimensional features or multiple views. Additionally, they plan to introduce Multilayer Perceptron (MLP) layers into the GCN framework to further reduce feature dimensionality and improve model efficiency and performance.

This research provides new ideas and methods for semi-supervised learning in multi-view data, offering significant theoretical and practical value. By introducing GCNs, the researchers successfully addressed the challenges of semi-supervised learning in non-graph data, laying the foundation for future related studies.