Existing approaches to the self-consistency exploration in Large Language Models (LLMs) primarily rely on post-hoc selection, overlooking the inherent logical structures essential for reasoning. To address this gap, we pioneer a new perspective by introducing S 2 CR , a self-supervised reasoning framework that presents the LLM consistency from internal modeling while coupling retrieval-augmented generation (RAG) for knowledge integration and process supervision. This framework operates in four stages: Information Retrieval patches the parametric knowledge of LLMs and provides factual grounding for evaluation; Response Generation generates multiple candidate responses for input to explore diverse reasoning paths; Consistency Evaluation quantifies logical consistency by aligning extracted triples from both the generated responses and retrieval information; and Duality Synergy Optimization (DSOP) further bolsters the consistency performance through two complementary modules, Introspection-Driven Self-correction Guidance (IDSG) and Fine-Grained Consensus Alignment (FGCA). Experiments conducted on three public datasets POPQA, Biography, and ALCE-ASQA demonstrate that S 2 CR achieves objective quantification of internal self-consistency and significantly improves performance ranging from 3.19% to 23.49% over Baseline ⋄ across diverse foundational LLMs, e.g. , GPT-3.5-turbo, GPT-4o, and open-source models LLaMA3-8B and LLaMA3-70B.

Existing trajectory representation learning methods are usually designed for specific downstream tasks, resulting in limited generality. Specifically, (1) current grid graph-based representation learning approaches do not simultaneously consider the contextual information and neighborhood features of nodes. Moreover, they fail to effectively represent the varying importance of nodes within their neighborhoods, inhibiting a comprehensive capture of the graph’s network structure. (2) Existing models exhibit poor performance in representing long trajectories, lacking in adequately depicting their long-term semantic features. To address these problems, we propose a novel trajectory representation learning model namely DGAN-TRL (Deep Graph Attention Network-based Trajectory Representation Learning). The model integrates Deepwalk and GAT algorithms, taking into account both the contextual information and neighborhood features of nodes, thereby capturing node features and the graph’s network structure more effectively. It then combines graph attention mechanisms with self-attention mechanisms to delve into the local features of trajectory sub-graphs, better representing the long-term semantic features of trajectories. Finally, the experimental results on two real-world trajectory datasets demonstrate that DGAN-TRL outperforms baselines in modeling short, medium, and long trajectories.

In recent years, large language models (LLMs) have achieved remarkable progress in dialogue generation. However, existing automatic evaluation methods still face challenges in diverse scenarios, particularly in terms of limited generalization and low alignment with human judgment. To address these issues, we propose a paraphrase-based evaluation framework that integrates Abstract Meaning Representation (AMR) with general-purpose language models (GPT). This approach generates diverse paraphrases across lexical, syntactic and stylistic dimensions to enhance the coverage of traditional evaluation metrics. Experimental results show that incorporating paraphrase augmentation significantly improves the correlation between automatic metrics and human evaluation on multiple datasets. Additionally, extensive experiments on six mainstream LLMs demonstrate the effectiveness and generalizability of the proposed method. This study offers new insights into improving the human alignment of automatic evaluation and lays a foundation for the application and optimization of LLMs in open-domain dialogue systems.

Knowledge-aware Recommendation (KGR) aims to utilize a knowledge graph to provide rich side information for items in a recommendation system and construct a unified graph containing users, items, and entities. In this paper, we present a new graph neural network for user-item-entity interaction modeling, named Graph Attention Intent Network, it employs different strategies to aggregate user and item information to generate high-quality representations. Typically, the description of user-item interactions is modeled as a bipartite graph, which overlooks the relations between users and between items, a significant aspect of realistic recommendation. Therefore, we propose a framework, named knowledge-aware multi-view contrastive learning for recommendation. It can explore effective user-user and item-item relations in the heterogeneous network of KGR, construct a user social graph and an item similarity graph, and combine the information of the two views into user-item-entity interaction modeling to enhance the representation of users and items. We introduce cross-graph contrastive learning to facilitate the integration of heterogeneous information while alleviating the sparse labeling problem of recommendation tasks. Experimental results on three benchmark datasets show that our model is more effective than other state-of-the-art models.

Fuzzy systems (FSs) have been widely utilized in diverse domains, such as pattern recognition, intelligent control, data mining, and bioinformatics due to their strong interpretation and learning abilities. Traditionally, FSs have mainly been applied to model Euclidean data. However, with the emergence of scenarios involving graph data, such as social networks and traffic route maps, which inherently possess non-Euclidean structures, there is a need to develop FS modeling methods suitable for graph data while retaining the advantages of traditional FSs. This article presents a novel FS called graph fuzzy system (GFS) specifically designed for modeling whole graph data. The concepts, modeling framework, and construction algorithms are systematically developed. First, the article defines GFS-related concepts, including the graph fuzzy rule base, graph fuzzy sets, and graph consequent processing unit (GCPU). Second, the learning framework for GFS is proposed. It includes a novel K-Means with graph similarity measure clustering approach (KM-GSM) for generating antecedents in GFS and a new consequent parameters learning algorithm based on graph neural network (GNN). Moreover, three different versions of the GFS implementation algorithm are developed and thoroughly evaluated through experiments on various graph prediction datasets. The results demonstrate that the proposed GFS inherits the advantages of mainstream GNNs methods and conventional FSs methods while achieving superior performance in whole graph prediction compared to existing approaches.

Graph data modeling is nontrivial due to the challenges to ensure model interpretability and handle data uncertainty. While methods derived from deep learning models, such as graph neural networks (GNNs), are able to handle graph data, the interpretability is limited. Graph fuzzy systems (GFSs) based on the fuzzy rules and fuzzy inference have been proposed to improve interpretability, but the existing methods are developed for whole graph prediction only and cannot deal with node prediction, which is a more common task in graph data modeling. To tackle the challenges, a novel GFS for node prediction (GFS-node) is investigated in this study. For this purpose, the concepts, framework, and algorithms of GFS-node are systematically developed. First, several related concepts are defined, including the node fuzzy rule base, node fuzzy set, and node consequent processing module (NCPM). A general framework for GFS-node is then presented, where the construction of antecedents and the consequents of fuzzy rules are analyzed. Furthermore, a concrete implementation method of GFS-node is designed. In particular, the kernelKvirtual central nodes clustering (KVCN) algorithm is proposed to develop the algorithm for antecedent generation, and the linear message passing network (LMPN) is adopted to develop the algorithm for consequent generation and learning. Experiments are carried out on multiple benchmark datasets, and the results show that GFS-node combines the advantages of both traditional fuzzy systems and classical GNNs for node prediction.

The emergence of new network attacks, especially multi-step attacks which exhibit complex patterns, presents challenges to network security. Intrusion detection in network traffic is one of the core means to ensure network operation security. Traditional intrusion detection methods identify abnormal traffic by modeling the patterns of normal traffic. This strategy focuses on the spatial characteristics of data, ignoring the temporal characteristics of traffic evolution, and may not fully capture the patterns of multi-step network attacks. To address this problem, we propose a Temporal-Gated Graph Neural Network (TGGNN) framework based on graph clustering sampling. Firstly, each network traffic data is viewed as a graph node. We use hierarchical graph clustering methods to sample representative points and further construct local evolutionary structures with their temporally adjacent nodes. Then, we design a novel way of constructing graph data that balances both local and global aspects of network traffic data. Based on the GGNN (Gated Graph Neural Network), we designed a Temporal-Gated Mechanism in the information propagation part, emphasizing the learning of the local evolutionary structure of the graph. Furthermore, a bidirectional LSTM network is used to further enhance the learning of dynamic patterns in network traffic data. Experimental results based on the UNSW-NB15 dataset demonstrate that our approach not only surpasses the performance of four recent baselines but also eliminates the need for feature engineering by adopting an end-to-end approach.

The gut microbiome has been regarded as one of the fundamental determinants regulating human health, and multi-omics data profiling has been increasingly utilized to bolster the deep understanding of this complex system. However, stemming from cost or other constraints, the integration of multi-omics often suffers from incomplete views, which poses a great challenge for the comprehensive analysis. In this work, a novel deep model named Incomplete Multi-Omics Variational Neural Networks (IMOVNN) is proposed for incomplete data integration, disease prediction application and biomarker identification. Benefiting from the information bottleneck and the marginal-to-joint distribution integration mechanism, the IMOVNN can learn the marginal latent representation of each individual omics and the joint latent representation for better disease prediction. Moreover, owing to the feature-selective layer predicated upon the concrete distribution, the model is interpretable and can identify the most relevant features. Experiments on inflammatory bowel disease multi-omics datasets demonstrate that our method outperforms several state-of-the-art methods for disease prediction. In addition, IMOVNN has identified significant biomarkers from multi-omics data sources.

Calligraphy has a long history as one of the Chinese outstanding traditional arts. However, calligraphy, as an artistic derivative of Chinese characters, suffers from a lack of teachers, a variety of disciplines, and confusing aesthetic standards. With the development of calligraphy, the need for calligraphy evaluation has also gradually increased, but the traditional way to evaluate calligraphy works relies too much on the work of calligraphy and ignores the motion of writing. To address such problems, we construct a multimodal dataset that contains images of calligraphy works, time series data of writing movements and aesthetic evaluation labels. To exploit the time series data of writing movement, we propose a strong benchmark that combines the Long Short-Term Memory network with the K-nearest neighbor algorithm. The proposed model achieved the best accuracy compared with baseline methods. And the evaluation results of our model are close to that of calligraphy experts. This study serves as a guide to the aesthetic evaluation of computational calligraphy.

From the perspective of philosophy, ontology relations denote ultimate semantic relations of related knowledge concepts. Beyond doubt, it is still a very difficult problem on how to automatically depict and construct ontology relations because of its high abstractness. Some latest research attempted to realize ontology relation learning by learning abstract hierarchies or similarities among knowledge concepts. Inspired by the requirements of associative semantic cognition like in the human brain, a constructivist ontology relation learning (CORL) method is put forward in this study by borrowing the idea of the constructivist learning theory. Wherein, two following points are supposed: 1) each symbol knowledge is looked as a token of representing certain abstract pattern and 2) each pattern denotes a type of relation structures on other patterns, or a directly observed event data, such as physical sensing data, natural image, sound data, text word etc. So, ontology relation could be considered as the associative support degrees from other knowledge concepts to the target concept, which reflects how one knowledge ontology can be demarcated by other knowledge concepts. Then, the knowledge network can be employed to represent an entire domain knowledge system. Meanwhile, an associative random walk mechanism (ARWM) on knowledge network can be considered to explain the semantic generative process of every document. Thus, CORL can be realized by integrating ARWM into an extended latent Dirichlet allocation (LDA) model. Some theoretical and experimental analysis are done. The corresponding results demonstrate that CORL can obtain effective associative semantic relations among concept words, and gain some novel characteristics in better representing knowledge ontology than existing methods.