Heterogeneous graph embedding has recently advanced with two primary directions: self-supervised learning to address the scarcity of labeled data and metapath-free approaches to eliminate reliance on handcrafted, domain-specific metapaths. While metapath-free methods capture structural and semantic information directly from neighbor relationships, their performance often remains unstable across datasets and varying label scales, as they still rely on semi-supervised training. Conversely, self-supervised learning techniques, such as contrastive learning and masked graph autoencoders, achieve remarkable results by extracting information intrinsically from the data. However, these approaches often depend on predefined metapaths, which limit their ability to generalize and constrain representation quality. This paper proposes Relation-aware Heterogeneous Masked Graph Autoencoder (RHMGA), a novel metapathfree self-supervised learning framework that bridges these gaps. Our method integrates a semantic relation network with masked graph autoencoders to jointly reconstruct node features and relation-specific adjacency matrices. By capturing both structural and semantic information during aggregation and reconstruction, RHMGA enables robust self-supervised pretraining without predefined metapaths. Extensive experiments demonstrate that our approach consistently outperforms state-of-the-art methods, achieving superior performance and adaptability across diverse heterogeneous graph datasets.

Despite considerable research into information diffusion, most models focus on static networks. Networks in the real world change over time and information is lost when a node disappears. We sought to resolve the problem of information loss by defining information and information loss in general terms. Experiments demonstrate the efficacy of the proposed scheme in resolving problems of information loss in real-world networks using only a few edges.

Most previous research on influence maximization has focused on static social networks, despite the dynamic nature of networks in the real world. The computational cost imposed by recalculating results in response to dynamic changes precludes the use of conventional updating algorithms when dealing with large-scale networks. In this study, we developed a novel approach to estimating the influence of nodes through the creation of Influence SubGraphs. We also developed methods by which to update Influence SubGraphs to overcome the problem of influence maximization in dynamic social networks. Experiment results demonstrated the efficacy of the proposed scheme, in achieving influence propagation performance comparable to that of state-of-the-art methods with far lower memory requirements and far shorter execution times.

Social network-based team formation problem has been widely studied from different aspects. However, the skills in earlier works were treated equally, and cannot be substituted by other related or similar skills. In addition, assigning experts who possess alternative skills for a required skill is not allowed. To better fit real world scenarios, we propose a novel hierarchical skill model to let skills interchangeable. By considering the communication cost and the personnel cost, we develop an optimization framework under the hierarchical skill model to deal with the trade-off between communication and personnel cost. The experiments show that our proposed framework and the hierarchical skill model is reasonable and has better performance than earlier works.

The disappearance of important nodes which are prominent characters in a social network may lead the social network to a broken structure. Many previous works have discussed reconstructing such networks using the network topology to devise an approach that finds a substitute node for a deleted node and generates appropriate links to avoid a fragmentation of the network. A common used property in finding substitute node is centrality, but calculating some kinds of centrality may spend too much time on re-scanning the graph. Thus, we propose a local approach, CLOMADE, standing for Choosing LOcal MAximum DEgree. We only need to scan the whole graph once for calculating degree. We choose a node with local maximum degree to be the substitute node and generate new links from the substitute node to other nodes. The experiments show that CLOMADE outperforms previous works in execution time.