ArchiveFeature generation (FG) aims to enhance the prediction potential of original data by constructing high-order feature combinations and removing redundant features. It is a key preprocessing step for tabular scientific data to improve downstream machine-learning model performance. Traditional methods face the following two challenges when dealing with the feature generation of scientific data: First, the effective construction of high-order feature combinations in scientific data necessitates profound and extensive domain-specific expertise. Secondly, as the order of feature combinations increases, the search space expands exponentially, imposing prohibitive human labor consumption. Advancements in the data-centric artificial intelligence (DCAI) paradigm have opened novel avenues for automating feature generation processes. Inspired by that, this paper revisits the conventional feature generation workflow and proposes the multi-agent feature generation (MAFG) framework. Specifically, in the iterative exploration stage, multi-agents will construct mathematical transformation equations collaboratively, synthesize and identify feature combinations exhibiting high information content, and leverage a reinforcement learning mechanism to evolve their strategies. Upon completing the exploration phase, MAFG integrates the large language models (LLMs) to interpretatively evaluate the generated features of each significant model performance breakthrough. Experimental results and case studies consistently demonstrate that MAFG framework effectively automates the feature generation process and significantly enhances various downstream scientific data mining tasks.
Driven by rapid advancements in data acquisition and storage technologies, vast amounts of unlabeled tabular data have been collected and stored across various industries. Clustering analysis serves as a fundamental method for uncovering the latent grouping structures within such data. Currently, traditional clustering algorithms still predominantly constitute the methods employed for processing this type of data. Deep learning techniques and large model technologies, known for their powerful representation and inference capabilities, are primarily employed for handling unstructured data types like images, text, and audio; their advantages remain largely underexploited in the context of structured tabular data processing. In 2025, Nature published TabPFN, a foundation model for tabular data capable of efficiently handling classification and regression tasks, offering a novel perspective for tabular data learning. Building upon TabPFN, this paper proposes a clustering method for tabular data based on pretrained foundation models with synthetic data. The method comprises two main phases: a pretraining phase and an iterative inference phase. The pretraining phase leverages traditional clustering algorithms and TabPFN model to obtain initial pseudo-labels for the unlabeled tabular data. The iterative inference phase then employs the fine-tuned TabPFN model to cyclically update these pseudo-labels until the final clustering results are obtained. Extensive experimental analyses conducted on benchmark datasets demonstrate that the proposed method significantly improves the clustering performance of seven representative clustering algorithms.
Efficient exercise recommendation is crucial in intelligent education to provide learners with personalized learning resources based on their cognitive states. Existing methods often rely on learners’ mastery levels of knowledge concepts, yet critically overlook contextual information embedded in response sequences data, thus limiting the enhancement of learners’ knowledge transfer abilities. This study introduces a context-aware adaptive exercise recommendation method for response sequences data (CAE-RRS). By constructing a learning context representation module that integrates knowledge concept and exercise contexts, the method offers a richer contextual representation. Furthermore, it incorporates a cognitive state representation module fused with temporal large language models (LLMs) and implements a context-aware adaptive exercise recommendation module based on knowledge graphs. This module dynamically links exercises, knowledge concepts, and learner characteristics, enabling precise and adaptive recommendations according to learning context features. It effectively addresses the challenge of context-aware adaptive recommendations, bolstering the model’s contextual perception and adaptability. Consequently, it ensures learning effectiveness while enhancing the novelty of the learning process and the efficiency and accuracy of exercise recommendation systems. Comparative and ablation experiments across the AAAI 2023 and NeurIPS_t34 datasets demonstrate that our model surpasses baseline models in both accuracy and novelty.
Test-time adaptation (TTA) seeks to dynamically recalibrate a deployed model through online fine-tuning with unlabeled or sparsely labeled data, addressing performance degradation caused by distribution shift, sensor noise, illumination changes, and other real-world challenges. It has demonstrated great promise in latency- and robustness-critical applications such as autonomous driving, remote healthcare, and video surveillance. However, existing multimodal TTA methods typically overlook the quality variance across views—low-quality or faulty views data can introduce harmful gradients—and lack mechanisms to preserve internal temporal consistency, undermining stability in dynamic scenarios. To tackle these issues, we propose CVPTA (confidence-guided view pruning and temporal contrastive attention), a unified framework comprising three core modules: 1) confidence-guided dynamic attention, which computes each view’s confidence via its predictive entropy and uses a softmax over inverted entropy scores to attenuate high-entropy views data during feature fusion; 2) view pruning, which evaluates image-modal quality using the Laplacian variance (for blur) and histogram skewness (for exposure), discards views below a quality threshold, and maintains a constant view count by interpolating adjacent high-quality views data or injecting Gaussian noise—dramatically reducing noise accumulation; 3) temporal contrastive self-supervision, which treats adjacent frames of the same view as positive pairs and non-adjacent frames as negatives, and applies a contrastive loss to enforce temporal feature consistency. We evaluate EVPTA on the public Kinetics-50-C and VGGSound-C perturbation benchmarks under the same online-update settings as prior work. Results show that EVPTA boosts Top-1 accuracy by approximately 2.3% and 0.7% on each benchmark and retains over 0.2% gains even under extreme noise conditions. Ablation studies further confirm the individual and synergistic contributions of all three modules. EVPTA requires no extra annotations, integrates seamlessly with existing multimodal systems, and delivers both efficiency and robustness—offering significant theoretical insights and practical value.
Text-based person retrieval, a pivotal task in multimodal intelligent surveillance systems, seeks to identify target pedestrian images from large-scale databases using free-form textual descriptions, with critical applications in public security and video forensics such as suspect tracking in criminal investigations and cross-camera forensic analysis. Conventional approaches predominantly operate under the idealized assumption of perfectly aligned image-text pairs, thereby overlooking the pervasive presence of complex noisy data—incorrect or ambiguous associations between visual instances and their textual annotations—that inevitably arise from human annotation biases, web-crawled data impurities, or semantic granularity mismatches between localized visual attributes and holistic textual contexts. To bridge this critical gap, we propose semantic-informed noisy correspondence learning, a unified framework that systematically addresses both noise identification and robust learning through dual innovative mechanisms. First, the semantic-aware noise identification criterion holistically integrates intra-modal semantic consistency with cross-modal interaction signals, enabling precise discrimination of false-positive and false-negative correspondences via adaptive thresholding. Second, the noise-robust complementary learning paradigm implements differentiated optimization strategies: positive learning with momentum contrastive alignment enforces discriminative feature learning on purified reliable pairs, while negative learning via entropy minimization constrains overfitting to noisy subsets. Extensive experiments on three public benchmarks validate the effectiveness of the proposed approach, achieving superior performance under both synthetic and real-world noise scenarios.
In crowd counting tasks, the high cost of data annotation severely limits the broad application of fully supervised methods. To significantly reduce reliance on labeled data, semi-supervised counting methods leveraging large amounts of unlabeled data have emerged as the mainstream research direction. However, existing semi-supervised approaches typically rely on iteratively generating pseudo-labels for training, and their performance is significantly constrained by two types of uncertainty factors: 1) Epistemic uncertainty (model’s inherent lack of knowledge), leading to unstable pseudo-label generation quality; and 2) Aleatoric uncertainty (inherent noise and ambiguity in the data), making the model susceptible to background interference and distribution bias. To address these challenges, we propose the progressive cognition-guided dual-domain network for semi-supervised crowd counting (PCDNet). It designs a progressive cognition-guided pseudo-label refinement mechanism which employs multi-granularity filtering to effectively eliminate low-quality pseudo-labels and significantly mitigate epistemic uncertainty. Furthermore, it introduces a spatial-frequency dual-domain joint counting loss. By enforcing the model to learn consistent and robust feature representations in both the spatial and frequency domains, this loss effectively constrains aleatoric uncertainty arising from data bias, enhancing the model’s generalization capability for complex scenes. Extensive experiments on four mainstream crowd counting datasets demonstrate that PCDNet significantly outperforms existing semi-supervised methods, with advantages being particularly pronounced when labeled data is scarce. The results fully validate the effectiveness of the proposed method in addressing epistemic and aleatoric uncertainty, generating high-quality pseudo-labels, and enhancing model robustness and generalization performance, providing a superior solution for semi-supervised crowd counting.
With the rapid development of computer vision, natural language processing, and deep learning technologies, multimodal vision-language representation learning models have demonstrated outstanding performance in tasks such as image captioning, text-to-image generation, and visual question answering. Such models have consequently become a focus of research in both academia and industry. However, the multimodal nature and complexity of these models provide attackers with more diverse attack vectors. Attackers can craft adversarial examples to mislead these models into outputting incorrect, harmful, or false content, posing an increasingly serious security threat to such models. This paper systematically reviews the current state of research on multimodal vision-language models. It also categorizes and summarizes the adversarial examples-based attack methods that have emerged in recent years against these models, as well as their defense strategies. Furthermore, it provides a detailed overview of the fundamental principles, implementation approaches, and conclusions of relevant studies. On this basis, this paper discusses the current state and future directions of security research in multimodal vision-language representation learning. Finally, the paper envisions the future prospects of applying vision-language representation learning technology in combination with interpretability techniques.
Time series are widely present in domains such as industry, finance, transportation, meteorology, and healthcare, where they hold significant analytical and practical value. However, missing values caused by human errors or unforeseen events often compromise the completeness of time series, thereby undermining the accuracy and reliability of subsequent analysis and decision-making. Incomplete time series typically exhibit complex temporal dependencies and inter-variable correlations, making effective imputation a challenging task. Deep learning methods, with their powerful modeling capabilities, have emerged as a promising solution to address these challenges and have become an active area of research. We provide a comprehensive review of incomplete time series imputation methods based on deep learning. The survey begins by introducing the relevant concepts and definitions, followed by a systematic categorization and analysis of representative methods, highlighting their core ideas and characteristics. Commonly used open-source datasets and evaluation metrics are then summarized, along with a well-designed experimental framework. The experiments evaluate 10 mainstream deep learning models in terms of imputation performance and efficiency under various conditions, including different missing scenarios, missing rates, variable dimensions, sequence lengths, and downstream task performance. We analyze the experimental results to derive comprehensive insights, then recommend appropriate imputation methods based on these findings. Finally, we discuss current research trends and future directions in this field, concluding the survey.
Privacy auditing is a crucial issue of data governance, aiming to detect whether data privacy has been protected effectively. Typically, scholars would protect personal private data to meet differential privacy guarantees by perturbing data or adding noise to them. Especially in scenarios of machine learning, an increasing number of differential privacy algorithms have emerged, claiming a relatively stringent level of privacy protection. Although rigorous mathematical proofs of privacy have been conducted before the algorithms’ release, the actual effect on privacy in practice is hardly assured. Due to the complexity of the theory of differential privacy, the correctness of their proofs may not have been thoroughly examined, and imperceptible errors may occur during programming. All of these can undermine the extent of privacy protection to the claimed degree, leaking additional privacy. To tackle this issue, privacy auditing for differential privacy algorithms has emerged. This technique aims to obtain the actual degree of privacy-preserving of differential privacy algorithms, facilitating the discovery of mistakes and improving existing differential privacy algorithms. We survey the scenarios and methods of privacy auditing, summarize the methods from three aspects―data construction, data measurement, and result quantification, and evaluate them through experiments. Finally, we present the challenges of privacy auditing and its future direction.
Data deduplication is a vital technology for efficiently managing big data, widely adopted in cloud storage systems to reduce redundancy and save space. To integrate deduplication with encryption, convergent encryption has become a common approach. This method allows for the encryption of data while still enabling deduplication to produce the same ciphertext for identical plaintexts. However, cloud service providers’ outsourcing models and the deterministic nature of convergent encryption can introduce data security issues. The encryption patterns of data can become predictable, potentially exposing sensitive information to attackers, which may create serious security implications. As a result, encrypted data deduplication has emerged as an important research topic in cloud storage security. We firstly introduce the concept of data deduplication, encrypted deduplication algorithms, and discuss the security challenges associated with encrypting and deduplicating data in cloud storage. Then we review the current research status from both attack and defense perspectives, covering three main types of attacks: brute force attacks, which try to decrypt data through extensive guessing; frequency analysis attacks, which exploit frequency characteristics in ciphertexts; and side-channel attacks, which leverage information from response or traffic characteristics. For each attack type, representative defense strategies are analyzed along with their strengths and weaknesses. Finally, we highlight the challenges faced by existing encrypted data deduplication defenses and suggest future research directions aimed at improving these techniques.
Malicious domain name detection is a critical component of network intrusion detection systems, enabling the rapid identification of network attacks through domain name requests. Machine learning methods overcome the limitations of blacklist mechanisms and improve detection accuracy. However, challenges such as the high variability of domain name structures and the complexity of real-world environments lead to low detection efficiency and poor robustness in practical applications. To address these issues, a malicious domain name detection technology based on domain name semantic graph learning is proposed, leveraging semantic graph association analysis for efficient detection. Specifically, 12 months of domain request data from China Science and Technology Network are first collected, encompassing 3.33 billion access records, including more than 6.5 million malicious domain name entries across 284 attack types. Semantic analysis reveals significant differentiation between domain categories, yet considerable feature overlap in certain regions degrades classifier performance. To tackle this, a domain association graph model based on character-level semantic similarity is proposed. By integrating features of neighboring domains, the model enhances semantic representations in overlapping regions, thereby improving detection performance. The method includes filtering noise characters through structural similarity analysis, constructing a dynamic domain semantic graph using an online aggregation algorithm, and training a multi-head attention-based message-passing graph model with node-degree-weighted samples. Finally, a multi-layer neural network classifier is employed for malicious domain detection. Experimental results demonstrate that the proposed method achieves an average precision rate of 0.96 and a recall rate of 0.97 on the dataset of different types of malicious domain names. Furthermore, the model exhibits strong online adaptability, achieving high detection rate and robustness.
Software vulnerabilities pose a serious threat to the safe and stable operation of computer systems and software, so the research related to their automatic detection has been receiving extensive attention. Unlike traditional static vulnerability detection tools that use predefined rules provided by human experts to analyze the code, graph neural network (GNN)-based vulnerability detection methods have surpassed the traditional methods in some datasets by automatically learning the vulnerable code patterns. However, in the current GNN-based vulnerability detection methods, the design of GNN model is not combined with the characteristics of the code itself, which leads to poor detection effect on the real vulnerability code dataset. In this paper, we propose a learnable hierarchical graph representation vulnerability detection method LHG-VD, which is characterized by proposing a learnable readout function for the limitation of the traditional readout function, and designing a cross-granularity loss function based on the idea of comparative learning for the problem of maintaining the local structural information of the code in the process of graph pooling. Experimental results on real vulnerability datasets show that the F1 value of LHG-VD is 71.5%, which is improved by 4.9% compared with DeepWukong, a slice-level detection method, and 8.9% compared with AMPLE, a function-level detection method.
Airborne systems serve a pivotal function in the aerospace industry. Their exceptional safety requirements make the formal verification of software requirements a critical and pressing issue. However, with the ever-increasing complexity of airborne system requirements and the growing number of onboard devices, formal verification has encountered the serious challenge of state space explosion. To alleviate this problem, this paper proposes a novel approach for the modeling and compositional verification of airborne system requirements based on time partitioning. This approach utilizes the time dimension to decompose a complex verification system into mutually independent components, enabling the independent verification of each component and subsequently synthesizing the overall verification results of the entire system. The feasibility and practicality of the proposed method are demonstrated through a real-world case study. The evaluation further shows that the proposed method not only enables the verification of systems that are beyond the capability of traditional monolithic verification approaches and effectively alleviates state space explosion, but also avoids false alarms that may arise from neglecting time-based partitioning. This innovative method provides a new and promising technical path for the formal verification of software requirements in airborne systems, contributing to enhanced verification accuracy and efficiency, and ultimately ensuring the operational safety and reliability of airborne systems.
Supercomputing has rapidly developed from traditional CPU clusters to heterogeneous platforms. With the type conversion of hardware platforms, it faces significant challenges in optimizing computing software programs and performance evaluation. Currently, some international mainstream parallel program performance analysis tools and software generally have low compatibility with domestic supercomputing heterogeneous system processor products, often requiring instrumentation and recompilation of code, and low accuracy in single node performance data collection. To improve these shortcomings, we propose a floating-point performance data collection method for heterogeneous system computing software. This method is based on the domestic supercomputing system verification platform to develop and verify the floating-point performance collection prototype. At present, effective collection of single node and multi node performance indicator data has been achieved, and it is non-invasive to the original program. There is no need to modify the code of the monitored program for monitoring in a plug-in manner, making it highly versatile. Finally, we conduct comparative experimental analysis with three types of programs: rocHPL, Cannon, and mixbench, and conduct performance data collection monitoring research on ResNet (residual network, ResNet) program for AI computing. We demonstrate that the collection method proposed in this article has high accuracy, achieves the expected collection effect in experiments, and has good reference value for program optimization, verifying the effectiveness of the proposed method.