Xunlian Wu

Xunlian Wu (仵训练) is a PhD candidate 🎓 in the School of Computer Science and Technology at Xidian University (XDU). During his master's studies, he was supervised by Prof. Ziyu Guan in the State Key Laboratory of Integrated Services Networks. In his PhD program, he is supervised by Prof. Peng Gang Sun at the Xi'an Key Laboratory of Big Data and Visual Intelligence. His research interests include graph neural networks, graph clustering, and community detection.

Email  /  Google Scholar  /  Twitter  /  Github  /  Researchgate

• [2024.03] ✌️ I received Research and Innovation Funds of Xidian University funding.
• [2023.12] ✌️ I won Outstanding Graduate Student.
• [2023.11] ✌️ I received Outstanding Doctoral Dissertation Funding.
• [2023.10] ✌️ I won Tencent Social Scholarship.
• [2023.10] ✌️ One paper has been accepted by ESWA.
• [2023.09] ✌️ One paper has been accepted by Physica A.
• [2022.12] ✌️ I won Outstanding Graduate Student.
• [2022.10] ✌️ I won Tencent Social Scholarship.
• [2022.09] ✌️ I won first-class scholarship for doctoral students.
• [2022.05] ✌️ One paper has been accepted by IEEE T-KDE.
• [2022.02] ✌️ One paper has been accepted by Physica A.
• [2021.09] ✌️ One paper has been accepted by IEEE T-PAMI.
• [2021.03] ✌️ One paper has been accepted by IEEE T-CYB.

More

My research interests mainly include social network analysis, self-supervised learning and graph neural networks. I'm also interested in their applications such as community detection, influence maximization, deep graph clustering,recommendation and bioinformatics. * denotes corresponding author. The selected papers are listed as follows:

The complexity of influence spread poses significant challenges to its accurate modeling. Traditional approaches, such as the Linear Threshold (LT) model, offer a simplified depiction of this phenomenon but overlook the temporal decay of influence. To address this limitation, we introduce a Dynamic Linear Threshold Model with Decay (DLTD), incorporating three distinct decay strategies: random decay, linear decay, and nonlinear decay. Comparative analyses on real-world datasets reveal that DLTD achieves superior accuracy over the LT model.

We propose an unsupervised method for community detection based on a Deep Dual Graph attention Auto-Encoder (DDGAE). For a graph, the DDGAE considers its attribute information and modularity information as two views, and learns a common representation based on a self-training process during which we can also determine the community assignment.

Identifying communities plays an essential role in disclosing modular structures with specific functions or properties in different kinds of complex networks. We develops a novel algorithm, Graph Embedding based on Motifaware Feature Propagation (GEMFP) for community detection. We conduct extensive experiments on 14 real-world datasets, and the results show that our algorithm tends to be more effective for community detection.

Nowadays, communities strongly overlap due to the fact that people strongly interact, which makes community detection even more challenging. The paper develops a new algorithm by rearranging ‘indivisible’ blocks(RaidB). We compare RaidB with the existing state of the art methods for non-overlapping and overlapping community detection. The results show that RaidB tends to achieve better performance especially on sparse networks with unobvious communities and networks with strongly overlapping communities.

Nowadays, boundaries between circles are more and more indistinct because people probably involve more than one circle. We develops an influence percolation method (IPM) for identifying overlapping communities. An interesting finding is that two nodes tend to be indivisible if one is a seed, influence percolates into the other exceeding a certain frequency, and this threshold is mainly determined by the networks’ density.

We are concerned with using user-tagged images to learn proper hashing functions for image retrieval. However, tagging data suffers from noises, vagueness and incompleteness. Different from previous unsupervised or supervised hashing learning, we propose a novel weakly-supervised deep hashing framework which consists of two stages: weakly-supervised pre-training and supervised fine-tuning. The framework is general and does not depend on specific deep hashing methods.

Most previous works assumed that each view has complete data. However, in real-world datasets, it is often the case that a view may contain some missing data, resulting in the problem of incomplete multiview clustering (IMC). We present the adversarial IMC (AIMC) framework. We explore two variants of AIMC: 1) AAIMC and 2) GAIMC with different strategies for learning multiview common representation. AAIMC tries to capture hidden information of the missing data by MDI via the elementwise reconstruction and the GAN. GAIMC directly searches the multiview comprehensive representation and infers the information of the missing views from the instances of the same cluster.