Community search over directed graphs is a fundamental graph problem, with applications in various areas such as social network analysis, financial fraud detection, and biology. A common requirement in real applications is to return a community with a constrained size. However, most existing studies primarily focus on the cohesiveness of the community and ignore the constraint on community size. Therefore, this paper investigates the problem of searching (k_c,k_f) -truss community with a size constraint over directed graphs (denoted by SCTC), which aims to find a subgraph that satisfies csup ≥ k_c and fsup ≥ k_f among all connected subgraphs that contain the query vertex q and have at least l and at most h vertices, where q, l, h, k_c , and k_f are specified by the query. We develop an exact solution for SCTC search by exploiting the properties of (k_c,k_f) -truss subgraphs. The approach first identifies the maximal (k_c,k_f) -truss subgraph to narrow the search space, and then performs a refinement search within it to efficiently support the SCTC search. We investigate two search strategies, expansion and peeling. The expansion strategy guides the search order using a score function and optimizes the search strategy based on clique, while the peeling strategy designs a level-based mechanism to remove low-support edges one level at a time iteratively. Extensive experimental results on real-world datasets show that our proposed algorithm improves both the quality of the resulting community and the search efficiency.
Active learning, which lets the model select the most informative data within a limited annotation budget for annotation, often achieves comparable performance with fewer labelled samples. However, despite its potential benefits, active learning has not been extensively explored in Named Entity Recognition (NER), as NER is considered more challenging than traditional tasks commonly addressed in the active learning literature. In this paper, we propose a novel deep active learning framework called Deep AUC Maximisation-based Active Learning (DAMAL), which learns uncertainty-aware representations for unlabeled data. Notably, DAMAL, due to its foundation in AUC maximisation, excels in addressing label sparsity, making it a well-suited approach for NER tasks. Experimental results across several NER tasks highlight the advantages of the proposed framework over existing active learning methods for NER.
In multi-agent systems, the inappropriate behavior of agents may lead to a series of negative side effects (NSE) on the environment or other agents, thereby affecting the overall performance of the system. Existing methods for mitigating negative side effects predominantly adopt homogeneous cooperation paradigms, where agents are assumed to have aligned objectives and behave cooperatively. However, these methods critically fail to account for multi-agent scenarios involving heterogeneous objective conflicts, where agents pursue divergent or even conflicting objectives that complicate coordination and side effect mitigation. To address this issue, we present a Dynamic Adjustment Lexicographic Multi-Objective (DALMO) reinforcement learning approach that dynamically adjusts the degree of side effects avoidance based on the strategies of other agents. This approach estimates the side effects of other agents through belief modeling to adjust its own policy and uses a multi-objective integration to balance individual rewards and overall system performance. To evaluate the effectiveness of our approach, we designed a multi-agent grid world experiment in which agents have conflicting objectives. We compared our approach with existing approaches and the experimental results show that our approach achieves higher overall rewards in different scenarios with objective conflicts. This demonstrates that our approach enhances the overall performance of multi-agent systems with objective conflicts by employing corresponding side effect mitigation strategies.
Controlling disease risk (diagnostic errors) poses a significant challenge in the field of medicine, largely due to the inherent data imbalance between affected (positive) and healthy (negative) individuals. To address this challenge, we propose an effective classification strategy tailored for disease risk control. Our research aims to regulate the false negative rate (type I errors) while minimizing the false positive rate (type II errors). We employ the Neyman-Pearson paradigm, combining oversampling and feature selection techniques to balance the training dataset, thereby improving disease risk prediction accuracy. To evaluate our approach, we conduct comprehensive numerical experiments on an imbalanced Heart Failure Clinical Records dataset. The results demonstrate that oversampling techniques and feature selection can enhance the accuracy of disease risk control to a certain degree, and our proposed strategy achieves robust and commendable performance.
Conventional neuron editing methods for language model privacy protection suffer from a critical limitation: uniform intervention strategies inadvertently exacerbate cross-category leakage via the privacy seesaw effect. Current methodologies overlook the functional divergence between category-specific neurons (activated solely by single-class private data) and cross-category neurons (responsive to multiple privacy classes). This study introduces a taxonomy-driven editing framework that identifies privacy neurons through integrated gradient attribution and categorizes them via multi-task activation patterns. Our method combines adversarial patching for category-specific neurons with class-conditional regularization for cross-category neurons. Evaluations on BERT and GPT-2 demonstrate that our method can effectively reduce privacy data exposure and leakage while maintaining model utility (perplexity increase <8
Histopathological image analysis is a critical task, as it provides the gold standard for clinical cancer diagnosis. Traditional histopathology image analysis relies on extensive datasets with annotations for training and testing purposes. However, the digitisation and labelling of histopathological images is a time-consuming and labour-intensive process, which limits the amount of annotated data. To alleviate this problem, we propose a novel Text-Guided Multi-Task Few-shot Learning framework called TGMT-FSL. Specifically, we first extract image and text feature information separately, then construct a feature map between the two, and finally construct an image-level similarity space and use multiple loss functions to train the model parameters. Extensive experiments evaluate the effectiveness of our methods on EBHI dataset, which fully demonstrate the potential of clinical application of the method in the early diagnosis and precision treatment of gastrointestinal cancer.
Computer-aided diagnosis is poised to image-based detection of molecular alterations, potentially improving the dilemma of time-consuming and costly clinical genetic testing. Prevalent approaches employ a two-stage scheme, which first identifies tumor regions from whole slide images (WSIs) and then exhaustively learns patches split from these areas with a patch-wise supervision model under WSI labels to characterize WSIs. However, parts of the patches are uninformative, and spatially close patches may exhibit similar discriminative power. Moreover, a model must traverse all possible magnifications to select the optimal one, as various cancer subtypes may necessitate different optimal recognition magnification. All these constraints lead to significant inefficiencies in the application of deep models on large histopathology images. To alleviate these limitations, this paper proposes a joint magnification and attention sampling-based cascade network for BRCA mutation detection from histopathology images. Specifically, to tackle redundant patches and tedious traversal, large regions from WSIs are downsampled to generate region-wise attention maps, which are the basis for selecting representative patches from different spatial locations at different magnifications. Meanwhile, to enhance the effectiveness of patch-level representations, the cascade feature fusion network (CFF) is proposed to integrate high-level representations from early stages to facilitate low-level feature learning in subsequent stages. Experimental results on the public ovarian dataset TCGA show our method can effectively reduce resource requirements and run 2.6–6.5 times faster in inference with a slight classification performance loss compared to the SOTA methods.
Mathematical Word Problem (MWP) is one of the tasks in natural language reasoning. MWP presents mathematical exercises as narratives, requiring the translation of textual descriptions into mathematical expressions. Unlike standard mathematical problems, MWP solvers must interpret and extract relevant information from the narrative. This process enables the evaluation of solvers across multiple skills, including reading comprehension, information identification, text-to-mathematics translation, and mathematical problem-solving. While reasoning performance did not follow a scaling law, Chain-of-Thought (CoT) has emerged to enable step-by-step reasoning, improving the performance of all reasoning tasks, including MWPs. However, most existing MWP datasets primarily focus on linguistic features such as lexical diversity, grammatical accuracy, or rely on augmenting already validated datasets through various data augmentation methods. As a result, they remain limited in terms of mathematical complexity and the diversity of the expressions presented. Consequently, examining the full potential of LLMs in tackling a broader array of mathematical concepts remains limited, and coverage for discovering hidden linguistic and reasoning deficiencies of LLMs is also limited. In this paper, to bridge these gaps, we propose generating MWP datasets by LLMs, incorporating a wider range of mathematical expressions and topics. The generated MWPs are followed by a manual verification process conducted by the researcher to ensure accuracy and quality. This approach enabled to generation of TEFMP, which contains 978 problems spanning around 13 diverse mathematical topics. Then LLMs are questioned to solve the generated problems and their variations. Through this, we reveal the weaknesses of current LLMs in solving math problems.
Hypergraphs have emerged as a critical technique to enhance recommendation performance by directly modeling global and higher-order associations in interaction graphs. However, existing hypergraph methods typically focus on high-order relations, neglecting the role of pairwise relations between nodes, and excessively entangle user and item features during convolution. Aiming to tackle the challenges, we introduce a new Hypergraph Disentangling and Cross-Level Contrastive Learning mechanism, named HDCCL. Specifically, we construct dual-perspective hypergraphs of users/items and disentangle them into three complementary views: clique graphs, star graphs, and dynamic weighted graphs, achieving dynamic multiscale feature fusion from both spatial and spectral perspectives. Additionally, a novel bidirectional attention-based feature decoupling mechanism is designed to differentiate domain-specific unique features from cross-domain commonalities, while explicitly integrating neighborhood aggregation signals from lower-order graphs to effectively enhance the feature entanglement issue. Furthermore, this paper introduces a cross-level contrastive learning module to constrain the semantic consistency of multi-views and enhance the robustness of representations. Comprehensive experiments on publicly accessible datasets demonstrate the superiority of our HDCCL.
Automatic Depression Detection (ADD) methods utilize multimodal data, including text, audio, and visual information, to facilitate early clinical diagnosis and intervention. However, current ADD approaches predominantly model global or single-scale features, thus inadequately capturing fine-grained local depressive cues and insufficiently exploiting complementary information across modalities. In this paper, we propose MSTDD, a Multi-Scale Transformer-based method designed to address these limitations by effectively extracting and integrating depression-related features at multiple scales. Specifically, MSTDD employs modality-specific multi-scale encoders to capture hierarchical local depressive indicators, and introduces a multimodal cross-attention fusion mechanism to promote robust feature interaction between modalities. Extensive comparative evaluations and ablation experiments conducted on two benchmark depression datasets-DAIC-WOZ (AVEC 2017) and E-DAIC (AVEC 2019)-demonstrate that MSTDD outperforms state-of-the-art ADD methods, achieving average F1-scores of 0.82 on DAIC-WOZ and 0.80 on E-DAIC. Additionally, we conduct experiments comparing baseline models under different fusion strategies, further validating the effectiveness of our proposed method.
Time series anomaly detection (TSAD) is a critical task in scientific domains like astronomy, where identifying rare and scientifically meaningful events is important. However, existing TSAD methods face limitations when applied to astronomical settings. Generic methods often exhibit high false positive rates due to their neglect of domain-specific semantics. Moreover, the lack of interpretability undermines scientific trust, and single-modality approaches fail to leverage the diverse sources of knowledge. To address these challenges, we propose MAVI, a Multimodal large language model-enhanced Anomaly Validator and Interpreter tailored for astronomical TSAD. Rather than serving as a primary detector, MAVI operates as a lightweight post-processing framework that filters and explains anomaly candidates produced by base TSAD methods. MAVI introduces two novel innovations: first, a multi-modal in-context learning strategy that retrieves numerical, visual, and textual anomaly templates to integrate multi-source knowledge; second, a domain-guided chain-of-thought prompting mechanism that emulates astronomers' analytical reasoning to enhance both accuracy and interpretability. Experiments on six astronomical datasets show that MAVI substantially reduces false positives while maintaining high recall, and provides expert-aligned, interpretable rationales for anomaly decisions.
Accurately locating security vulnerabilities in ever-growing codebases is critical for protecting modern software ecosystems. Yet real-world detectors must contend with three intertwined hurdles: extreme class imbalance, thousand-token functions that exceed Transformer limits, and vulnerability patterns that span semantic, contextual, and syntactic dimensions. We propose Sparse-MoE, a syntax-aware, multi-view mixture-of-experts framework that addresses these obstacles within a single architecture. The pipeline (i) splits ultra-long functions along abstract-syntax-tree boundaries to preserve long-range semantics without quadratic attention, (ii) encodes each block through complementary semantic (SBERT), contextual (CODE2VEC), and syntactic (CodeBERT-C) views, and (iii) routes embeddings to a small set of lightweight experts via a risk-aware sparse gate while enforcing load balance. Training combines cross-entropy with block-level InfoNCE alignment and an expert-usage regularizer. Sparse-MoE attains F1 = 0.94, Precision = 0.97 on the highly imbalanced BigVul benchmark and 0.70 accuracy on the balanced Devign dataset, while analysing 2 k-line functions with just 1.24GB GPU memory and 13 ms latency per inference. These results demonstrate that syntax-aware splitting, risk-driven sparse gating, and multi-view contrastive alignment jointly yield an efficient and scalable solution for large-scale vulnerability screening.
Service-oriented knowledge base (KB) has been applied to a variety of applications. In online platforms, the official accounts registered by companies display service mentions (e.g., train ticket booking) to provide services for users. To facilitate the downstream tasks, we propose to construct a dynamic service KB. Specifically, the service KB contains accounts in different service domains. Moreover, since these service mentions for the same service have different names, the service KB should store the canonicalized service mentions for each account. Despite the fruitiness, existing KB construction and NLP approaches rely on abundant annotated labels. However, in real-world applications, it is impractical and expensive to annotate abundant account and service mention labels, due to a large number of accounts and services. In this paper, we propose an end-to-end dynamic service KB construction system. First, we infer the service domain of each account under limited labeled accounts. Second, without service mention labels, we propose to adaptively separate service mentions into disjoint partitions, where service mentions inside the same partition provide the same service. Third, our system can be updated with dynamic scenarios naturally, including service changes, new accounts, and new service domains. The experiments on real-world datasets and a large-scale online A/B testing demonstrate the effectiveness and high practicality of our system.
Objective assessment of melasma is a critical indicator for evaluating treatment efficacy. However, current clinical practice primarily relies on subjective physician judgment for assessment, which results in significant inter-rater variability. Establishing a reliable automated segmentation method for melasma is essential to achieve precise assessment, and this requires support from high-quality datasets. The Melasma Image Dataset (MEMI-DS) was constructed to address this limitation, comprising 716 images with corresponding annotation files, for training melasma segmentation models. To validate the effectiveness of the dataset, 10 segmentation models (including UNet, UNet++, DeepLabV3, HRNet, DeepLabV3+, LinkNet, MANet, PAN, TransUet, and SegFormer) were systematically evaluated. Experimental results demonstrated that MEMI-DS can effectively distinguish the segmentation performance of different models. Notably, the segmentation performance of models represented by SegFormer was significantly improved by applying data augmentation techniques. These findings not only confirm the effectiveness and practicality of the dataset but also provide reliable technical support for the objective assessment of melasma.
Natural Language Processing (NLP) workflows in biomedical domains face unique challenges due to specialized terminologies and the need for high precision in downstream applications. This study presents a systematic framework for preprocessing and analyzing biomedical texts, with a focus on evaluating tokenization strategies and their impact on representation learning. We have proposed a dual-phase approach: first, benchmarking various tokenizers across efficiency and domain-specific accuracy metrics; second, integrating context-aware embedding techniques to enhance semantic capture. Our experiments reveal that SciSpacy outperforms conventional tokenizers in biomedical term recognition despite computational trade-offs, while custom-trained BPE models achieve a 22
Recent advances in large language models (LLMs) are revolutionizing recommender systems (RS). LLM-based RS typically relies on a large number of training samples (more than 10
In clinical scenarios, radiology reports are crucial for radiologists in disease diagnosis, while writing reports manually is laborious and subjective. Existing radiology report generation (RRG) approaches primarily introduce external clinical knowledge to improve the generation quality of radiology reports, but they overlook the inherent ranking information among the injected knowledge elements. In this study, we propose a rank-aware framework named RankRRG, which incorporates a ranking loss to strengthen the model’s representation learning capability. It encourages the model to learn the relative ordering of semantically similar reports through similarity-based supervision, enabling the model to learn more discriminative and clinically meaningful features. Experiments conducted on the widely-used Chest X-ray benchmark MIMIC-CXR demonstrate that our proposed RankRRG achieves competitive performance compared with state-of-the-art methods, specifically surpassing the second top-tier with a margin of 0.7 F_1 score, highlighting the effectiveness of incorporating ranking information into RRG.
Current image captioning methods mainly adopt the autoregressive framework that operates through a next-token prediction paradigm. A non-autoregressive method called diffusion models has shown superiority in image generation. However, their potential in image captioning remains underexplored due to the visual-language misalignment. In this work, we present a novel Masked Conditional Diffusion model (MC-Diffusion). It contains a discrete denoising diffusion probabilistic model (D3PM) and a pre-trained vector quantized variational autoencoder (VQ-VAE). Specifically, we first extract discrete image features via VQ-VAE. Conditioned on these discrete image features, the discrete diffusion model generates captions through transformer blocks to establish discrete-to-discrete alignment. Furthermore, we propose a simple yet effective guidance method, named Masked Condition Strategy (MCS). Compared with classifier-free guidance, our proposed method achieves finer-grained visual-language alignment while demonstrating superior capability in model guidance. Experiments on the CUB-200 dataset show that the proposed method performs better than baselines on several metrics. Compared with classifier-free guidance, MCS achieves similar performance on reference-based metrics (e.g., BLEU, Meteor, etc.) while alleviating the hurt on CLIPScore.
Knowledge Tracing (KT) aims to optimize teaching effectiveness by assessing students' knowledge states. Existing KT methods primarily rely on learning knowledge states from historical interaction sequences but often fail to fully utilize the dataset, particularly neglecting the importance of problem embedding in KT. Although some models attempt to improve problem embeddings, they still struggle to effectively capture and model the complex relationships between problems and skills, limiting their ability to reveal the underlying information and dependencies between them. To address this challenge, we propose a Dual Hypergraph-based Question Embedding Model with Multiple Relations for Knowledge Tracing (DHGE). Specifically, we design an Explicit Relation Hypergraph (ERH) module and an Implicit Relation Hypergraph (IRH) module to capture the intricate relationships between problems and skills from both local and global perspectives. The explicit module constructs hyperedges based on annotated problem-skill mappings for local information propagation, while the implicit module captures global dependencies by establishing connections between each problem-skill pair and utilizing adaptive weights to dynamically learn problem-skill associations. Furthermore, we introduce a Question Embedding Fusion (QEF) module that integrates both local and global features of problem embeddings to generate high-quality representations. Extensive experiments demonstrate that our model outperforms existing methods, highlighting the effectiveness of high-quality problem embeddings in knowledge tracing.
