Link prediction aims to predict the likelihood of connections between node pairs, which is an important task in modern graph data analysis. Recently, graph neural networks have been widely applied and have achieved great success in link prediction. However, methods based on standard graph neural networks often struggle to capture fine-grained relational signals because of their node-centric design, which aggregates neighborhood information independently and fails to explicitly capture the joint structural dependencies (e.g., common neighbors) relative to the target link. To address this issue, we propose a novel link prediction method, IGLP (Importance-Guided Link Prediction), which takes a link-centric perspective. IGLP employs a unified importance scoring framework that adaptively quantifies the relevance of each neighbor to a target link by integrating both structural and feature information. Based on this framework, we present a concrete implementation of IGLP with modules of Adaptive Neighborhood Aggregation (ANA) and Common Neighborhood Awareness (CNA). The ANA module selectively aggregates informative neighbor features by jointly considering connectivity patterns and feature similarities, while the CNA module encodes structural similarity through neighborhood overlap. Extensive experiments are conducted on a series of datasets to validate the effectiveness of IGLP. The results demonstrate that IGLP achieves superior performance against state-of-the-art counterparts, particularly on sparse graphs, validating its ability to adaptively construct expressive link representations. This work offers a flexible and generalizable approach to advance GNN-based link prediction.

Graph self-supervised learning (GSSL) has shown great promise in addressing the label scarcity problem in real-world graphs, yet existing generative self-supervised learning (SSL) methods based on feature reconstruction still suffer from two key limitations: they typically underuse topological information and rely on a single, fine-grained reconstruction objective that fails to capture the multilevel semantic structure of graphs. These issues make the learned representations sensitive to the quality of node features and limit their ability to model complex graph semantics. To tackle these challenges, we propose a multiscale masked graph autoencoder framework, MMGAE, which unifies feature reconstruction with hierarchical semantic prediction. MMGAE first constructs an encoding tree by minimizing structural entropy on an enhanced graph, and then employs a multilayer graph neural network (GNN) encoder to learn node representations at different receptive-field scales. The final-layer representations are used for masked feature reconstruction, while intermediate-layer representations are supervised to predict multilevel communities defined by the encoding tree, enabling the model to jointly capture topology-aware and multiscale semantic information. Extensive experiments on both node classification and graph classification benchmarks demonstrate that MMGAE consistently achieves competitive or superior performance compared to state-of-the-art generative and contrastive baselines. These results highlight the effectiveness of combining structural entropy–driven encoding trees with multiscale masked graph autoencoders (MAEs), and showcase the potential of MMGAE as a general-purpose representation learner on graphs.

Network embedding has been widely investigated to learn low dimensional nodes representation of networks, and serves for many downstream machine learning tasks. Previous network embedding studies mainly focus on static networks, and cannot adapt well to the characteristics of dynamic networks which are evolving over time. Some works on dynamic network embedding have tried to improve the computation efficiency for incremental updates of embedding vectors, while others have made efforts to utilize temporal information to enhance the quality of embedding vectors. However, few existing works can fulfill both efficiency and quality requirements. In this article, a novel dynamic network embedding model named TPANE (Temporal Path Adjacency Matrix based Network Embedding) is proposed. It employs a new network proximity measure: Temporal Path Adjacency, which is capable of capturing the temporal dependency between edges as well as being incrementally computed in an efficient way. It evaluates the similarity between nodes via the count of temporal paths between them, rather than making random sampling approximation, and adopts matrix factorization to obtain embedding vectors. Link prediction experiments on various real-world dynamic networks have been conducted to show the superior performance of TPANE against other state-of-the-art methods. Time consumption analysis also shows that TPANE is more efficient in incremental updates.

Supervised link prediction aims at finding missing links in a network by learning directly from the data suitable criteria for classifying link types into existent or non-existent. Recently, along this line, subgraph-based methods learning a function that maps subgraph patterns to link existence have witnessed great successes. However, these approaches still have drawbacks. First, the construction of the subgraph relies on an arbitrary nodes selection, often ineffective. Second, the inability of such approaches to evaluate adaptively nodes importance reduces flexibility in nodes features aggregation, an important step in subgraph classification. To address these issues, a novel graph-classification based link-prediction model is proposed: Attention and Re-weighting based subgraph Classification for Link prediction (ARCLink). ARCLink first extracts a subgraph around the two nodes whose link should be predicted, by network reweighting, i.e. attributing a weight in the range 0-1 to all links of the original network, and then learns a function to map the subgraph to a continuous vector for classification, thus revealing the nature (non-existence/existence) of the unknown link. For leaning the mapping function, ARCLink generates a vector representation of the extracted subgraph by hierarchically aggregating nodes features according to nodes importance. In contrast to previous studies that either fully ignore or use fixed schemes to compute nodes importance, ARCLink instead learns nodes importance adaptively by employing attention mechanism. Through extensive experiments, ARCLink was validated on a series of real-world networks against state-of-the-art link prediction methods, consistently demonstrating its superior performances

Link prediction, which centers on whether or not a pair of nodes is likely to be connected, is a fundamental problem in complex network analysis. Network-embedding-based link prediction has shown strong performance and robustness in previous studies on complex networks, recommendation systems, and knowledge graphs. This approach has certain drawbacks, however; namely, the hierarchical structure of a subgraph is ignored and the importance of different nodes is not distinguished. In this study, we established the Subgraph Hierarchy Feature Fusion (SHFF) model for link prediction. To probe the existence of links between node pairs, the SHFF first extracts a subgraph around the two nodes and learns a function to map the subgraph to a vector for subsequent classification. This reveals any link between the two target nodes. The SHFF learns a function to obtain a representation of the extracted subgraph by hierarchically aggregating the features of nodes in that subgraph, which is accomplished by grouping nodes with similar structures and assigning different importance to the nodes during the feature fusion process. We compared the proposed model against other state-of-the-art link-prediction methods on a wide range of data sets to find that it consistently outperforms them.

MOTIVATION: The reliable and reproducible identification of gene interaction networks represents one of the grand challenges of both modern molecular biology and computational sciences. Approaches based on careful collection of literature data and network topological analysis, applied to unicellular organisms, have proven to offer results applicable to medical therapies. However, when little a priori knowledge is available, other approaches, not relying so strongly on previous literature, must be used. We propose here a novel algorithm (based on ordinary differential equations) able to infer the interactions occurring among genes, starting from gene expression steady state data. RESULTS: The algorithm was first validated on synthetic and real benchmarks. It was then applied to the reconstruction of the core of the amino acids metabolism in Bifidobacterium longum, an essential, yet poorly known player in the human gut intestinal microbiome, known to be related to the onset of important diseases, such as metabolic syndromes. Our results show how computational approaches can offer effective tools for applications with the identification of potential new biological information. AVAILABILITY: The software is available at www.bioconductor.org and at www.picb.ac.cn/ClinicalGenomicNTW/temp2.html.