Imperfect Information Learning Team(Talk by Wan Seng)

This is an online seminar. Registration is required.
【Imperfect Information Learning Team】
【Date】2023/January/6(Fri) 10:00-11:00(JST)

【Speaker】Wan Sheng, Nanjing University of Science and Technology

Title: Semi-Supervised Learning with Graph Contrastive Learning and Its Applications

Abstract: With the rapid development of Internet technology, there exist a large number of unlabeled data in many practical tasks of the real world. Since data labeling requires a lot of manpower and material resources, it is usually expensive to obtain labels of the data. The generalization ability of supervised learning using only a small number of labeled data is limited, and the performance of unsupervised learning relying on unlabeled data is also poor. To solve this problem, semi-supervised learning is applied to improve the learning performance by using a large number of unlabeled examples when there are few labeled examples.
As one of the most important methods of semi-supervised learning, graph-based semi-supervised learning has aroused extensive research interests by its solid mathematical foundation and ease of implementation. In recent years, Graph Neural Network (GNN) stands out from different types of graph-based methods and has been proved to have strong representation abilities. Nevertheless, when the number of the labeled examples is quite small, the limited supervision information is insufficient to train a GNN model with good representation ability. Therefore, we hope to extract additional supervision information from the graph data themselves to guide the model training process. In contrastive learning, where no label information is available for model training, representation learning can be performed via contrasting similar and dissimilar examples. Therefore, we aim to design different types of GNN models under the framework of contrastive learning for semi-supervised learning. More specifically, by investigating the characteristics of graph data, we aim to enhance the performance of contrastive GNN models in semi-supervised learning from various aspects, including node features and graph topology.### ### ****