Energy efficiency has become a critical requirement, driving recommendation systems for resource-constrained environments such as edge devices. Model quantization offers an effective way to build low-bitwidth models while preserving accuracy. However, user–item interaction graphs contain numerous nodes and complex topological structures, leading nodes to exhibit unique similarities and differences. Existing quantization methods uniformly process parameters in high-dimensional DNN layers (e.g., linear, convolutional, or attention layers), while inadequately capturing such similarities among node embeddings. This paper proposes GraphQ, a chunk-wise quantization framework for graph collaborative filtering that supports both the training and post-training phases in a unified perspective. Our core idea is to adaptively partition node embeddings into multiple chunks based on the distribution of embedding values, and then apply chunk-wise quantization. Specifically, for quantization-aware training (QAT), we introduce learnable low-precision quantization factors that partition node embeddings into multiple chunks and are dynamically updated following message passing. For post-training quantization (PTQ), we first cluster nodes and then partition their dimensions into chunks for weight clipping. Extensive experiments on four real-world datasets show that GraphQ outperforms state-of-the-art QAT methods by an average of 27.49% in Recall@10 under the 256-dimensional embedding and 2-bit settings, and surpasses PTQ methods by 78.64% on average under 4-bit settings.

Next point-of-interest (POI) recommendation aims to predict the next interested POI to the user based on their historical check-in data in location-based social services. Most existing studies have attempted to model user visiting behaviors via sequence-based and graph-based models, and have achieved impressive performance. However, there is still room to explore the implicit transition preferences and contextual multiple semantic relationships among various POIs. To this end, we propose a novel graph-based Multi-Relational Variational Contrastive Learning (MRVCL) method for next POI recommendation, which captures local similarity associations and global contextual dependencies among POIs. Specifically, an order-free local-relational adaptive weighting module is designed to alleviate the problem of insufficient utilization of multi-hop neighbor information caused by the limit of neighbor order of nodes. We then develop a contextaware global-relational encoding module to capture implicit semantic sequential relationships. Finally, generative variational-contrastive learning is employed to construct a continuous representation of the latent features and reinforce the quality of representation learning. Extensive experiments on two real-world datasets validate that our MRVCL outperforms existing state-of-the-art methods on various evaluation metrics. To facilitate future research, our code and data are open-sourced at https://github.com/LinCH-en/MRVCL.

Multimodal sarcasm detection aims to identify whether utterances express sarcastic intentions contrary to their literal meaning based on multimodal information. However, existing methods fail to explore the model’s "ability to understand" the semantics expressed by sentences in the image context from semantic diversity perspectives. In this paper, we propose a multi-view semantic awareness method, which concretizes semantics from multiple perspectives to improve the model’s ability to capture different semantic features. Specifically, two learnable prefixes are attached to the text representation respectively to construct semantic representations from both the literal meaning and sarcastic intention perspectives. Then, image-text information is further fused through cross-attention to guide the semantic representation of different perspectives in the image context. Finally, the semantics expressed by prefixes are strengthened through KL divergence, thereby encouraging the model to capture two distinctive semantic features. Experiments on benchmark datasets demonstrate the effectiveness of our method.

The increasing complexity of cyber threats and the emergence of new attack technologies have brought huge challenges to attack incident analysis and source tracing. Using cyber threat intelligence to build a Cyber security Knowledge Graph (CKG) provides a new technical solution for attack attribution. Constructing a CKG requires numerous entity and relationship triples extracted from unstructured cyber threat intelligence texts. However, existing entity and relationship joint extraction methods in cyber threat intelligence face two problems. Firstly, they share the same word embeddings for both subtasks, ignoring the fine-grained semantic differences between the subtasks. Secondly, they rarely consider the interaction between the features of the two subtasks, which is vital for capturing the semantic dependencies between the tasks. To address these issues, we propose a joint entity and relationship extraction model specifically designed for network security concepts. We utilize two lightweight Fourier networks with independent weights to build a feature extraction module for encoding fine-grained features for entity recognition and relationship extraction tasks. Furthermore, we use a subtask feature interaction strategy assisted by a gated attention mechanism to enhance feature interaction between entity recognition and relationship extraction tasks. Use fine-grained entity recognition task information to guide relationship extraction to capture semantic dependencies between tasks. Experimental results on a cyber threat intelligence dataset demonstrate that our model outperforms existing baselines.

Joint extraction of entity and relation is crucial in information extraction, aiming to extract all relation triples from unstructured text. However, current joint extraction methods face two main issues. Firstly, they rarely consider the semantic information of entity and relation labels, leading to models that fail to fully understand and utilize the rich semantics in these labels, thereby limiting their performance. Secondly, although table-filling methods are widely used, they focus only on the start or end positions and ignore deep interactions between tables, relying solely on word-level information. To address these issues, we propose the P-MHE framework based on prompt learning and multi-channel heterogeneous graph enhancement. First, we use prompt templates to construct semantic nodes for entity and relation type labels, initializing them along with words as nodes in a heterogeneous graph. We iteratively fuse these semantic nodes through a message-passing mechanism to obtain node representations suitable for entity and relation extraction tasks. Secondly, we design a multi-channel heterogeneous graph to model node relationships from different perspectives, enhancing feature interactions among different types of nodes. Finally, we aggregate the semantic node information of entity and relation type labels after iteration, constructing separate decoding tables for each entity and relation type to better adapt to their respective characteristics. We evaluated our model on four public datasets. Experimental results show that P-MHE outperforms existing models on multiple public datasets. Extensive additional experiments further validate the effectiveness of our model.

In the field of cybersecurity, analyzing malicious software or programs is crucial for preventing network attacks. Malicious code often exists in a stripped binary form to thwart analysis, presenting challenges for analysts. This study investigates inferring function names from stripped binary to aid security researchers in analyzing malicious code. We propose PTGFI, a Prompt-based Two-stage Generative framework for Function name Inference. The PTGFI framework transforms the task of inferring function names into a two-stage semantic generation problem. By capturing function descriptions of assembly functions and introducing prompt learning, effective inference of function names is achieved. In experiments, PTGFI outperforms the state-of-the-art model by 2.96 % in precision. Moreover, ablation studies demonstrate the effectiveness of advanced components within the PTGFI framework. We further validate the utility and reliability of function names generated by the PTGFI framework through case studies.

Communities usually exhibit similar opinions, similar functions, or similar purposes, and recent years have witnessed the the resurgence of community detection in various fields. The node attribute network has gradually become the mainstream of the community network, however, most of existing community detection methods are limited by the high-dimensional node attribute and network topology, leading to the suboptimal performance. Inspired by recent deep learning-based community detection methods, we focus on building an unsupervised deep learning architecture to handle high-dimensional data in complex networks for community detection. To this end, we propose a novel Community Detection method based Deep Clustering and Graph Convolution auto-encoder Network (CD-DCGCN). Our CD-DCGCN designs an end-to-end framework consisting of dual auto-operations, one is a graph convolution auto-encoder, the other is community auto-detection. In addition, we realize the cooperative work of the dual auto-operations by constructing a joint optimized function. Our experimental results on nine attribute network benchmark datasets show that the proposed CD-DCGCN can obtain promising performance compared with several popular baseline methods.

Bundle recommendation suggests a set of items to users against their complex needs, where user-bundle interaction learning is key. It is observed that Gromov’s δ-hyperbolicity of the interaction graph in bundle recommendation is smaller (lower is more hyperbolic) than those in traditional item recommendation when measuring a graph’s tree likeness. However, state-of-the-art bundle recommendation methods learn to embed the entities (user, bundle, item) of tree-like interaction graph in Euclidean space, which could cause severe distortion problems. We argue hyperbolic space provides a promising way to get accurate entity embeddings, with this paper proposing a novel bundle recommendation model. The model learns user preferences via hyperbolic graph convolution, aiming at decreasing the distortion of bundle graph node embeddings. Extensive empirical experiments conducted on two real-world datasets confirm that our model achieves promising performance compared to baseline methods representing state-of-the-art bundle recommendation methods.

Tag recommendation in software information sites is a significant task to help developers make distinctions among software objects. Most existing methods usually ignore the semantic information of code snippets in software information sites. To tackle this issue, we regard the code as a semantic enhancement signal, and propose a novel Code-Enhanced fine-grained semantic matching method for Tag Recommendation in software information sites (CETR) to learn the matching score between tags and software objects. In our CETR, code-enhanced semantic interaction is designed to capture fine-grained semantic relevance between tags and software objects. Specifically, code snippets and text descriptions in software objects are first encoded in two specialized ways to capture semantic information from the perspective of natural language and structural language respectively. The semantic interaction is then learned between code snippets and text descriptions. Besides, a hierarchy-aware semantic learning block attentively combines the features from different layers as the final feature to capture semantic information of different levels. Lastly, a pairwise sentence matching computation block is exploited to predict matching probability. Experimental results on four software information site datasets have demonstrated the effectiveness of our proposed CETR for tag recommendation compared with the state-of-the-art methods.

Deep clustering aims to cluster unlabeled real-world samples by mining deep feature representation. Most of existing methods remain challenging when handling high -dimensional data and simultaneously exploring the complementarity of deep feature representation and clustering. In this paper, we propose a novel Deep Variational Attention Encoder-decoder for Clustering (DVAEC). Our DVAEC improves the representation learning ability by fusing variational attention. Specifically, we design a feature-aware automatic clustering module to mitigate the unreliability of similarity calculation and guide network learning. Besides, to further boost the performance of deep clustering from a global perspective, we define a joint optimization objective to promote feature representation learning and automatic clustering synergistically. Extensive experimental results show the promising performance achieved by our DVAEC on six datasets comparing with several popular baseline clustering methods.

With the rapid growth of user scale, it is so meaningful to explore the users who carry more valuable information in the user group. The importance of the individual users in the recommender system can improve the recommendation efficiency of the recommender system, enhance the robustness of the recommender system, but there is little research work in this path and can not determine a more effective method. To solve this problem, we propose a method based on multi-label clustering to determine core user, and define the concept of correlation between user and label cluster, user location weight, considered the potential relationship between users and labels. At the same time, we proposed a new method based on multi-label clustering and core user, according to experiments demonstrate the effectiveness of the proposed algorithm, we have a more significant upgrade in the recommendation accuracy and diversity.