Large Language Models (LLMs) perform excellently in fake news detection tasks, but their outputs are often accompanied by hallucinations, i.e., generated content that is contradictory to facts. Previous studies have mostly mitigated hallucinations through prompt design. However, this paper reveals that regions in news articles which easily induce hallucinations in LLMs correspond closely to the most challenging regions for fake news detectors. In this paper, we propose a fake news detection framework (PHPFND) based on post-hoc processing of LLMs hallucination. Specifically, our framework includes a hallucination detection module (ISHD) based on information structuring that detects three types of hallucinations in LLMs in a targeted manner, and a hallucination-driven feature enhancement mechanism (HDFE) that incorporates hallucination signals as explicit features into sentence-level encoding and feature fusion to guide the model’s attention toward high-risk regions. Experimental results on two mainstream fake news datasets show that our proposed method significantly outperforms LLM-based baselines.

While short-video platforms have democratized news consumption, they simultaneously exacerbate the proliferation of misinformation. Despite the progress of recent cross-modal frameworks, two critical challenges remain unaddressed: task-irrelevant noise (e.g., background music) that creates spurious correlations and the limitations of existing pairwise alignment methods that fail to capture the holistic structural conflicts within the video-audio-text triad. To address these challenges, we propose a novel framework for short-video fake news detection based on Semantic Purification and Geometric Routing (SPGR). First, to mitigate noise, we introduce Semantic Disentanglement Masking (SDM) as an adaptive filter to suppress environmental noise. Second, to transcend reductionist pairwise views, we introduce an Audio-Adaptive Projector that models the multimodal triad as a parallelotope, using Gramian geometric volume to quantify holistic consistency. Crucially, by utilizing the purified masks and geometric volume as hierarchical cues, SPGR orchestrates a progressive “Intuition-then-Reasoning” workflow: samples first undergo structural conflict checks by “intuition” experts before being passed to “reasoning” experts for fine-grained semantic scrutiny. Experiments on FakeSV and FakeTT demonstrate that SPGR outperforms recent state of-the-art methods in accuracy. The code is available at: https://github.com/sjsgmm/SPGR.

Multi-modal learning has become a transformative approach in recommendation systems, leveraging diverse data types—such as visual, textual, and audio signals—to construct rich and comprehensive user preference profiles. Despite significant progress, existing methods often struggle with key challenges, including imbalanced data utilization, diverse inter-modal correlations, and task-specific variability, which limit their ability to fully exploit inter-modal relationships. To address these issues, we propose KANM$^{2}$L (KAN Enhanced Multi-modal Learning for Recommendation), a novel framework that integrates the strengths of the Kolmogorov-Arnold Network (KAN) with multi-modal learning. Specifically, KANM$^{2}$L: (1) introduces a KAN-enhanced dilated attention mechanism to effectively capture high-dimensional, complex visual dependencies, enabling scalable and efficient processing of intricate datasets; (2) employs a multi-modal adversarial network to align and fuse features across modalities, ensuring seamless integration and improved recommendation accuracy; and (3) incorporates a rotational loss function to stabilize and refine visual feature embeddings, leveraging historical interaction data for more consistent performance. Extensive experiments on real-world datasets demonstrate that KANM$^{2}$L achieves state-of-the-art performance, with improvements of up to 11.7% over existing methods. These findings underscore the potential of KANM$^{2}$L to advance the field of multi-modal recommendation systems by overcoming critical limitations and delivering robust, scalable performance across diverse recommendation tasks.

Embodied intelligence, which enables agents to interact with the physical world, has gained significant attention for its real-world applications. However, these systems face two key challenges: optimizing task instructions and addressing vulnerabilities to adversarial attacks. We propose an AI agent integrating a large language model (LLM), a CLIP model, and two contextual databases. Additionally, we introduce an adversarial attack framework that manipulates prompt attention by replacing high-attention words and appending adversarial suffixes, inducing hazardous behaviors while maintaining visual and semantic plausibility. Our work advances instruction optimization and highlights vulnerabilities in embodied intelligence systems. Experimental results show that our optimization approach generates precise, real-world instructions. At the same time, our adversarial attack achieves higher success rates with reduced computational overhead, advanced optimization, and robustness in embodied intelligence systems.

The utilization of automated detection of fake news videos significantly improves early intervention for misinformation on short video platforms. A shortcoming of existing approaches is their inability to fuse multimodal features effectively. They only perform simple fusion of multimodal features to predict fake news, without deeply considering inter-modality relations. Inspired by the way people read news videos, we propose a novel multimodal deep fusion network (MFSVFND) that can effectively capture the important details of unimodality in detecting fake news videos by learning inter-dependencies among multimodal features. We conducted extensive experiments on two large-scale dataset of fake news videos in different languages, and our model outperforms state-of-the-art methods, demonstrating the superiority of MFSVFND in detecting fake news on short video platforms.

Urban region profiling is essential for forecasting and decision-making in dynamic and noisy urban environments. However, existing approaches struggle with adversarial attacks, data incompleteness, and security vulnerabilities, which undermine predictive accuracy and reliability. This paper introduces Enhanced Urban Region Profiling with Adversarial Self-Supervised Learning (EUPAS), a robust framework that integrates adversarial contrastive learning with self-supervised and supervised objectives. To fortify resilience against adversarial attacks and noisy data, we introduce perturbation augmentation, a trickster generator, and a deviation copy generator, which collectively enhance the robustness of learned embeddings. EUPAS significantly outperforms state-of-the-art models in forecasting tasks, including crime prediction, check-in prediction, and land use classification, achieving up to 12.2% improvement in forecasting performance. Additionally, our model demonstrates superior resilience against transfer-based black-box and white-box attacks compared to baseline models. By addressing key security challenges in data-driven urban modeling, EUPAS provides a scalable and adversarially robust solution for smart city applications.

With the rise of social media, the rapid spread of rumors online has resulted in numerous negative effects on society and the economy. The methods for rumor detection have attracted great interest from both academia and industry. Given the widespread effectiveness of contrastive learning, many graph contrastive learning models for rumor detection have been proposed by using the event propagation structure as graph data. However, the existing contrastive models usually treat the propagation structure of other events similar to the anchor events as negative samples. While this design choice allows for discriminative learning, on the other hand, it also inevitably pushes apart semantically similar samples and, thus, degrades model performance. In this article, we propose a novel propagation fusion model called propagation structure fusion model based on node-level contrastive learning (PFNC) for rumor detection based on node-level contrastive learning. PFNC first obtains three augmented propagation structures by masking the text of each node in the propagation structure randomly and perturbing some edges in the propagation structure based on the importance of edges. Then, PFNC applies the node-level contrastive learning method between every two augmented propagation structures to prevent the samples with similar propagation structure from far away. Finally, a convolutional neural network (CNN)-based model is proposed to capture the relevant information that is consistent and supplementary among three augmented propagation structures by regarding the propagation structure of the event as a color picture, three augmented propagation structures as color channels, and each node as a pixel. The experimental results on real datasets show that the PFNC significantly outperforms the state-of-the-art models for rumor detection.

Molecular property prediction plays a crucial role in drug screening and discovery scenarios. The critical task of it is to obtain the embedding of effective molecular structures. Textual sequences and graphs are commonly used to describe molecules. Previous efforts have attempted to combine these modalities to address the issue of information loss in single-modal representations across diverse tasks. Therefore, integrating chemical information from different modalities should be considered for more accurate representations. Given the advantages of Transformers in various fields of artificial intelligence, leveraging the attention mechanism to integrate molecular sequence and graph representations is desirable for achieving improved molecular embeddings. To this end, we propose a deep learning model called MJAF and design a novel information fusion strategy based on joint attention mechanisms. This approach effectively harnesses the strengths of both molecular representation modalities, significantly enhancing the efficiency of embedding molecules. We conducted multiple experiments comparing our model with state-of-the-art methods, experimental results on 4 independent datasets demonstrate significant advancements achieved by our proposed model.

The rise of social media has posed a challenging problem of effectively identifying rumors. With the great success of contrastive learning in many fields, many contrastive learning models for rumor detection have been proposed. However, existing models usually use the propagation structure of other events as negative samples and regard more similar samples to anchor events as hard ones across all the training processes, resulting in undesirably pushing away the samples of the same class. Thus, we propose a novel contrastive learning model (CRFB) to solve the above problem. Specifically, we employ contrastive learning between two augmented propagation structure and fit a two-component (true-false) beta mixture model (BMM) to measure the probability of negative samples being true. In addition, we propose a CNN-based model to capture the consistent and complementary information between two augmented propagation structure. The experimental results on public datasets demonstrate that our CRFB outperforms the existing state-of-the-art models for rumor detection.

The complexity of the graph structure poses a challenge for graph representation learning. Contrastive learning offers a straightforward and efficient unsupervised framework for graph representation learning. It achieves unsupervised learning by augmenting the original views and comparing them with the augmented views. Several methods based on this framework have achieved significant progress in the field of graph representation learning. Despite its success, the factors contributing to good augmented views in graph contrast learning have received less attention. In order to address this issue, we introduce the graph info-min principle. We investigate the relationship between mutual information (MI) and good augmented views through experimental and theoretical analysis. Additionally, we present a new contrastive learning method called Info-min Contrastive Learning (IMCL). Specifically, The method comprises an adaptive graph augmentation generator and a pseudo-label generator. The graph augmentation generator ensures sufficient differentiation between the augmented and original views. The pseudo-label generator generates pseudo-labels as supervision signals, ensuring consistency between the classification results of augmented views and original views. Our method demonstrates excellent performance through extensive experimental results on various datasets.

\beginabstract Graph neural networks (GNNs) have shown strong performance in graph-based analysis tasks. Despite their remarkable success, the inherent homophilic message-passing mechanism (MP) makes GNNs challenging to generalize to heterophilic graphs. In addition, the MP explicitly exploits the connection relationships between local neighbor nodes making GNNs unable to maintain stable performance in the face of adversarial perturbation attacks. In this paper, we propose a new method to explore graph structure by removing explicit message-passing mechanisms and present a network embedding framework AMCNE with Adaptive Multi-hop Contrast loss (AMCLoss) to address these challenges. AMCNE only relies on a simple autoencoder to obtain node representations for classification and uses elaborate contrastive loss to drive nodes capturing complex structural information on heterophilic graphs. The comprehensive experiments show that AMCNE outperforms state-of-the-art baseline models on homophilic and heterophilic graphs and is more robust in the node classification task. \endabstract

Graph embedding is becoming increasingly popular due to its ability of representing large-scale graph data by mapping nodes to low-dimensional space. Current research usually focuses on transductive learning, which aims to generates fixed node embeddings by training the whole graph. However, dynamic graph changes constantly with new node additions and interactions. Unlike transductive learning, inductive learning attempts to dynamically generate node embeddings over time even for unseen nodes, which is more suitable for real-world applications. Therefore, we propose an inductive dynamic graph embedding method called AGLI by aggregating global and local influences. We propose an aggregator function that integrates global influence with local influence to generate node embeddings at any time. We conduct extensive experiments on several real-world datasets and compare AGLI with several state-of-the-art baseline methods on various tasks. The experimental results show that AGLI achieves better performance than the state-of-the-art baseline methods.

Due to the rapid spread of fake news on social media, society and economy have been negatively affected in many ways. How to effectively identify fake news is a challenging problem that has received great attention from academic and industry. Existing deep learning methods for fake news detection require a large amount of labeled data to train the model, but obtaining labeled data is a time-consuming and labor-intensive process. To extract useful information from a large amount of unlabeled data, some contrastive learning methods for fake news detection are proposed. However, existing contrastive learning methods only randomly sample negative samples at different training stages, resulting in the role of negative samples not being fully played. Intuitively, increasing the contrastive difficulty of negative samples gradually in a way similar to human learning will contribute to improve the performance of the model. Inspired by the idea of curriculum learning, we propose a curriculum contrastive model (CCFD) for fake news detection which automatically select and train negative samples with different difficulty at different training stages. Furthermore, we also propose three new augmentation methods which consider the importance of edges and node attributes in the propagation structure to obtain more effective positive samples. The experimental results on three public datasets show that our model CCFD outperforms the existing state-of-the-art models for fake news detection.

In recent years, graph neural network (GNN) has become the most important method for graph representation learning. However, most GNNs focus on using the message passing mechanism to guide the information aggregation between neighbors, which results in the over-smoothing and weak robustness. To address the above issues, we propose a novel graph representation learning framework via Adaptive Multi-layer Neighborhood Diffusion Contrast, called AM-NDC in this paper. Without using the message passing mechanism, AM-NDC can still capture the complex structural information between nodes through a neighborhood diffusion contrast loss. Experimental results show that AM-NDC outperforms the existing state-of-the-art models in both node classification and robustness against adversarial attacks. Our dataset and code are available at https://github.com/YJ199804/AM-NDC.

Network representation learning has attracted more and more attention, whether in academia or industry. It can convert large scale network data into low dimensional node embedding for various tasks, such as node classification, link prediction, etc. Recently, representation learning on dynamic networks has emerged, which is more similar to the situation in real life. Inspired by the interaction between entities and entities in real life, we propose a model to embedding temporal network via interest vectors (ETIV). The interest vector of an entity can be generated according to the historical entities of the interaction. Interest vector can infer the possibility of interaction between entities. Then, we can obtain the node representation by optimizing the possibility of interaction. For calculating the interest vector of a node, we use a multi-head attention mechanism to capture the information of the historical interaction nodes from different aspects. Moreover, according to the interaction time of historical nodes, we introduce a learnable time parameter to simulate the forgetting of historical information. We conduct experiments on three real data sets and find that our model performs better than state-of-the-art methods in various tasks.

A major challenge in the field of object detection is to design a model with high precision and high speed, which is necessary in real applications. In this paper, we study the problem of object detection, and propose a multi-layer feature fusion model for object detection, called MLFF. MLFF not only takes into account the rich location information and detail information of low-level features, but also considers the rich spatial information of high-level features. In MLFF, we develop three different components, Backbone, Neck and Head, to gradually extract and fuse low-level feature maps and high-level feature maps. In addition, CBAM attention mechanism is also used in feature fusion, which enables us to achieve more accurate feature extraction. Extensive experiments on PASCAL VOC2007 and VOC2012 datasets show that our MLFF model not only has the speed of one-stage object detection model but also has the detection precision of two-stage object detection model. Considering both precision and speed, MLFF outperforms state-of-the-art detection models significantly.

Network representation learning aims to generate an embedding for each node in a network, which facilitates downstream machine learning tasks such as node classification and link prediction. Current work mainly focuses on transductive network representation learning, i.e. generating fixed node embeddings, which is not suitable for real-world applications. Therefore, we propose a new inductive network representation learning method called MNCI by mining neighborhood and community influences in temporal networks. We propose an aggregator function that integrates neighborhood influence with community influence to generate node embeddings at any time. We conduct extensive experiments on several real-world datasets and compare MNCI with several state-of-the-art baseline methods on various tasks, including node classification and network visualization. The experimental results show that MNCI achieves better performance than baselines.

Network representation learning (NRL) is playing an important role in network analysis, aiming to represent complex network more concisely by transforming nodes into low-dimensional vectors. However, most of the current work only uses network structure and node attribute to learn network representation, and often ignores the historical interactions between nodes that will affect the future interactions. Therefore, we propose a network representation learning algorithm based on neighborhood influence sequence (NIS), by investigating the influence of node historical interactions on future interactions. We propose three kinds of influence when two nodes interact, and integrate them into NIS by introducing the Hawkes process. In experiments, we compare our model with existing NRL models on four real-world datasets. Experimental results demonstrate that the embedding learned from the proposed NIS model achieve better performance than state-of-the-art methods in various tasks including node classification, link prediction, and network visualization.

Due to the fast development of the Internet technology, a variety of social applications have emerged. The users of these social applications are expanding, forming a large-scale social network. In complex and huge social networks, there are a lot of valuable information waiting to be utilized, so that service providers can provide better services for users. On the other hand, to solve the challenge of information overload brought with the communication technology, the recommendation system is becoming more and more important in many fields. In the context of the social network, a recommendation system can better discover the relationship between users and preferences, thus giving a more accurate recommendation. However, most of the traditional recommendation systems mainly focus on providing personalized recommendation service to a single user which are not suitable for group recommendation such as recommendation for a team or a family. In this paper, we proposed a new method based on trust relationship in social network and attention mechanism of deep learning to solve the group recommendation problem. Specifically, we use attention mechanism to capture different group members's preference weight and we consider trust relation among group members when learning group preference in order to capture more information. Through this method, the accuracy of the recommended results can be further improved. The experimental results on Film Trust dataset show that the proposed algorithm in this paper is able to improve the quality of recommendation compared with the existing group recommendation methods.

With the wide usage of many electronic devices such as the mobile telephones, the social activities in the mobile terminal have increased greatly. The influence maximization problem is to find a small set of nodes in a mobile social network such that they can make to maximize the spread of influence under the certain models. Existing methods of influence maximization only consider information dissemination, without considering the effectiveness of the transmission probabilities between nodes. We need to ensure the effectiveness of the transmission probabilities to achieve the maximal influence in real applications. We design a new propagation model and propose an effective greedy algorithm for this model. On this basis, we further improve the algorithmic efficiency with dynamic pruning strategy and propose a new estimation technique to accelerate influence evaluation. We demonstrate the effectiveness and efficiency of our algorithms on the two real social networks.

MOTIVATION: Identification of disease-associated miRNAs (disease miRNAs) is critical for understanding disease etiology and pathogenesis. Since miRNAs exert their functions by regulating the expression of their target mRNAs, several methods based on the target genes were proposed to predict disease miRNA candidates. They achieved only limited success as they all suffered from the high false-positive rate of target prediction results. Alternatively, other prediction methods were based on the observation that miRNAs with similar functions tend to be associated with similar diseases and vice versa. The methods exploited the information about miRNAs and diseases, including the functional similarities between miRNAs, the similarities between diseases, and the associations between miRNAs and diseases. However, how to integrate the multiple kinds of information completely and consider the biological characteristic of disease miRNAs is a challenging problem. RESULTS: We constructed a bilayer network to represent the complex relationships among miRNAs, among diseases and between miRNAs and diseases. We proposed a non-negative matrix factorization based method to rank, so as to predict, the disease miRNA candidates. The method integrated the miRNA functional similarity, the disease similarity and the miRNA-disease associations seamlessly, which exploited the complex relationships within the bilayer network and the consensus relationship between multiple kinds of information. Considering the correlation between the candidates related to various diseases, it predicted their respective candidates for all the diseases simultaneously. In addition, the sparseness characteristic of disease miRNAs was introduced to generate more reliable prediction model that excludes those noisy candidates. The results on 15 common diseases showed a superior performance of the new method for not only well-characterized diseases but also new ones. A detailed case study on breast neoplasms, colorectal neoplasms, lung neoplasms and 32 other diseases demonstrated the ability of the method for discovering potential disease miRNAs. AVAILABILITY AND IMPLEMENTATION: The web service for the new method and the list of predicted candidates for all the diseases are available at http://www.bioinfolab.top. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

We investigate the problem of summarizing frequent subgraphs by a smaller set of representative patterns. We show that some special graph patterns, called δ-jump patterns in this paper, must be representative patterns. Based on the fact, we devise two algorithms, RP-FP and RP-GD, to mine a representative set that summarizes frequent subgraphs. RP-FP derives a representative set from frequent closed subgraphs, whereas RP-GD mines a representative set from graph databases directly. Three novel heuristic strategies, Last-Succeed-First-Check, Reverse-Path-Trace, and Nephew-Representative-Based-Cover, are proposed to further improve the efficiency of RP-GD. RP-FP can provide a tight ratio bound but has heavy computation cost. RP-GD cannot provide a ratio bound guarantee but is more efficient than RP-FP. We also make use of the similarity between sibling branches in the graph pattern space to devise another much more efficient algorithm, RP-Leap, for mining a representative set that can approximately summarize frequent subgraphs. Our extensive experiments on both real and synthetic data sets verify the summarization quality and efficiency of our algorithms. To further demonstrate the interestingness of representative patterns, we study an application of representative patterns to classification. We demonstrate that the classification accuracy achieved by representative pattern-based model is no less than that achieved by closed graph pattern-based model.

Several efficient frequent subgraph mining algorithms have been recently proposed. However, the number of frequent graph patterns generated by these graph mining algorithms may be too large to be effectively explored by users, especially when the support threshold is low. In this paper, we propose to summarize frequent graph patterns by a much smaller number of representative graph patterns. Several novel concepts such as delta-covergraph,jumpvalueanddelta-jumppatternare proposed for efficiently summarizing frequent graph patterns. Based on the fact that all delta-jumppatternsmust be representative graph patterns, we propose two efficient algorithms for summarizing frequent graph patterns, RP-FP and RP-GD. The RP-FP algorithm computes representative graph patterns from a set of closed frequent graph patterns, whereas the RP-GD algorithm directly mines representative graph patterns from graph databases. Experimental results show that RP-FP and RP-GD are able to obtain compact summarization in both real and synthetic graph databases. When the number of closed graph patterns is very large, RP-GD is much more efficient than RP-FP, while achieving comparable summarization quality.