Recently, source code summarization has become a promising area in software engineering research. Most previous works only focus on a single code property(e.g., token or syntax) and ignore cross-granularity interactions, limiting the quality of code summarization. To address these issues, we present TriGraSum, a tri-level graph-based architecture for source code summarization that leverages multi-granularity representations and an explicit node-mapping relationship among the three granularities to capture richer features of source code. Specifically, we use sub-token embeddings obtained from a pre-trained LLM to capture fine-grained lexical and naming information. We design a Syntactic-Level Local Graph (SLG) to efficiently extract essential syntactic patterns and provide a skeleton for multi-granularity integration, and a Statement-Level Key-Structure Graph (SKG) to extract core statement-level logic relationships. The SLG and SKG representations are encoded using Graph Attention Networks (GATs) and integrated through a hierarchical pyramid attention mechanism to achieve comprehensive feature fusion. Finally, the fused representations are utilized for code summarization. Experiments on the FunCom and EMSE-DeepCom datasets demonstrate that TriGraSum achieves state-of-the-art performance, with up to 1.6% BLEU improvement over the strongest baseline on EMSE-DeepCom dataset. This work lays a foundation for more accurate and interpretable code summarization.

Binary code similarity detection primarily aims to compare two or more binary code files or code snippets to identify their differences. Currently, binary code similarity detection is extensively applied in information security domains such as vulnerability detection, malware analysis, and code clone detection. However, existing binary code embedding methods rely on large amounts of labeled data and suffer from poor scalability; even with contrastive learning they cannot perform adaptive augmentation tailored to the specific characteristics of each CFG, and the resulting embeddings fail to effectively counteract perturbations introduced by different compiler versions and optimization levels. To address this challenge, our paper proposes a novel program-level binary code embedding method using Gravity-Model Augmented Graph Contrastive Learning. Specifically, we introduce a graph augmentation strategy tailored for Control Flow Graphs (CFG), leveraging the concept of “gravity” between nodes and their intrinsic “mass” to dynamically compute probabilities for control flow dropping and feature masking. The gravity model can effectively simulate the impact of different compiler versions and optimization levels on the control flow graph of binary programs. This strategy generates two augmented views for effective contrastive learning. We implemented and evaluated the BEGA method using multiple open-source datasets, including BinaryCorp-3M, Coreutils. Experimental results show that, on program-level binary code similarity tasks, BEGA attains higher accuracy than existing graph contrastive learning methods under both cross-optimization and cross-version settings, and further exhibits advantages over NLP-based methods in certain cross-optimization scenarios. This work presents a robust and efficient unsupervised learning paradigm for program-level binary code similarity detection.

Heterogeneous graphs model complex scenarios and are increasingly important in representation learning. Traditional meta-path-based models first decompose the graph into several subgraphs according to distinct meta-paths, then aggregate neighbor information within each subgraph, and finally integrate the aggregated information. However, this process may introduce redundancy, and many models fail to differentiate between various neighbor types.However, multiple meta-paths can introduce redundant edges and many models fail to discriminate between neighbor types, degrading performance. We introduce a multi-relational graph-based algorithm for heterogeneous graphs that merges meta-path subgraphs to minimize edge redundancy and employs a relational discriminator to filter out mismatched neighbor information. This improves handling of heterogeneous neighbors by preserving their structural and attribute diversity. Our comparative experiments on node classification and clustering across three public datasets confirm the algorithm’s effectiveness.

As an emerging self-supervised graph representation learning model, graph contrastive learning has garnered significant attention. These models do not rely on manual annotations but instead leverage unsupervised contrastive learning strategies to maximize the consistency of similar graph or node representations, thereby learning high-quality graph representations. To fully exploit the structural information within graphs, we propose a graph contrastive learning model, GCL-CCSE. The model combines community detection and subgraph auxiliary models to better capture both local and global structural information during the learning process. Specifically, GCL-CCSE not only considers node-level information in the contrastive learning process but also incorporates community and subgraph-level information, avoiding comparisons solely between nodes and the entire graph, thus preserving the uniqueness of node embeddings. The model performs data augmentation through random edge dropping and feature masking, encodes using a Graph Convolutional Network (GCN), and generates positive and negative samples through community partitioning and subgraph sampling. Experiments conducted on four datasets for node classification tasks demonstrate that GCL-CCSE outperforms existing graph representation learning algorithms across multiple metrics. Additionally, ablation experiments verify the contributions of community and subgraph information to model performance, while sensitivity experiments explore the impact of subgraph size and the number of communities on model performance. This research provides new insights into graph contrastive learning and significantly enhances the performance of the model in downstream tasks.

Data augmentation has been widely used across various research domains in recent years. However, data augmentation applied to real-world graph-structured data tends to suffer from weak information, such as missing structural elements, incomplete features, and limited label availability. Graphs with weak information lack adequate training data to guide the graph learning process, resulting in performance loss. This article presents a novel automated data augmentation for graphs with weak information, namely AutoDAW, which aims to derive more informative node representations from incomplete input graphs in an adaptive way. To learn local features at the subgraph level, we adopt a reinforcement learning method to optimize the aggregation range for each node. To capture global semantic information, we present a graph diffusion-based long-range propagation method to facilitate more effective node feature propagation and aggregation. The dual-stage information aggregation promotes synergistic interactions between the two node representations. Extensive experiments on real-world datasets demonstrate the effectiveness, generalization, and robustness of AutoDAW.

Abstract Community detection is one of the most popular researches in a variety of complex systems, ranging from biology to sociology. In recent years, there’s an increasing focus on the rapid development of more complicated networks, namely multilayer networks. Communities in a single-layer network are groups of nodes that are more strongly connected among themselves than the others, while in multilayer networks, a group of well-connected nodes are shared in multiple layers. Most traditional algorithms can rarely perform well on a multilayer network without modifications. Thus, in this paper, we offer overall comparisons of existing works and analyze several representative algorithms, providing a comprehensive understanding of community detection methods in multilayer networks. The comparison results indicate that the promoting of algorithm efficiency and the extending for general multilayer networks are also expected in the forthcoming studies.