Exploiting Graph Structures in Semi-Supervised Learning to Boost Classification Performance with Sparse Labels

Semi-supervised learning’s major limitations are evident when trying to address sparse labeled data especially in complicated classifications. For this reason, within the current work we consider a new approach with the help of which we can improve the classification outcomes in SSL scenarios with a limited availability of labeled samples, while using the data structures in the form of graphs. Our approach integrates key components: accurate construction of graph representations, increase in graph descriptors, Graph Neural Networks (GNNs), and label spreading. In our method, graph representations are employed to generalize the assignments for the relativity between samples; in this sense, the labels of the positive samples are spread to the negative ones. Feature augmentation together with multi-layer GNN architecture, enhances the extent of capability of the model in learning the data characteristics. Thus, in order to test the suitability of the proposed method, we carried out experiments on several benchmark datasets and compared it with other baseline methods. The findings indicate improvement of classification metrics, namely accuracy, precision, recall, and F1 score. Further, we compared approaches to constructing the input graph and augmenting features with the results showing our strategy as the best. Examples of the use of the methodology in text classification, image recognition and medical data analysis indicate the usability and resilience of the proposed approach. This work contributes to the development of SSL by addressing the issues of sparse labels and enhancing the classification accuracy thus have potential use in future studies.