Rule-based question answering aims to produce interpretable answers that satisfy explicit rule constraints. However, existing methods either depend on semantic matching and overlook the logical function of rules, or adopt end-to-end generation without rigorous reasoning, which makes it difficult to achieve both interpretability and robustness. To address these issues, we propose VGRP-MCTS, an explicit rule-guided interactive reasoning framework that combines rule selection with Monte Carlo Tree Search. The framework decomposes questions into rule-constrained executable actions and constructs a minimal yet sufficient evidence chain. Extensive experiments on three benchmark datasets show that our method significantly outperforms state-of-the-art baselines, yielding Micro-F1 improvements of 4.25% for answers, 0.35% for supporting rules, and 6.65% for their joint metric.

To address the susceptibility of existing MTS(multivariate time series)anomaly detection models to training set contamination and their limited ability to capture complex spatial-temporal correlations in MTS,a novel anomaly detection model based on spatial-temporal normalizing flow was proposed.This model employed the conditional normalizing flow to estimate the density of patterns in MTS,enabling robust anomaly detection even in the presence of contaminated training data.Additionally,a patched long short-term memory module was introduced to effectively learn long-term temporal dependencies within MTS,and a dynamic graph learning module based on attention mechanisms was devised to model the evolving correlations among different dimensions of MTS.Experimental results on three real-world cyber-physical system datasets demonstrate that the proposed model significantly outperforms state-of-the-art baselines in both detection accuracy and robustness.

Federated learning facilitates big data analysis and applications while safeguarding user privacy, thus emerging as a pivotal paradigm in distributed machine learning. Building on this foundation, multi-objective federated learning (MOFL) — which focuses on the collaborative optimization of multiple objectives such as communication cost and computational efficiency — has become one of the current research hotspots. However, MOFL still suffers from performance degradation in scenarios with heterogeneous data and dynamic network topologies, limiting its practical applicability. To address these challenges, this paper proposes a multi-objective hierarchical aggregation optimization method tailored for dynamic network structures. Specifically, a hierarchical aggregation mechanism is adopted to tackle the dynamic variations in client-side neural network models, which optimizes the training process of MOFL and significantly enhances computational efficiency under dynamic network and heterogeneous data environments. Experimental results verify that the proposed method achieves remarkable performance improvements across different data distributions: it attains an average performance enhancement of 50.73% compared with the NSGA-III algorithm. Furthermore, comprehensive comparative experiments with other state-of-the-art multi-objective optimization algorithms demonstrate its overall superior performance, confirming the scalability of the proposed method in practical scenarios.

Addressing edge-level data heterogeneity in hierarchical federated learning remains a major research challenge, as it often induces global model bias, slower convergence, and client unfairness that are difficult to reconcile. In this paper, we propose G-M3CS, a gradient-based multi-metric multi-objective client selection algorithm, along with its enhanced version G-M3CS+. By integrating the NSGA-II algorithm, we formulate optimization objectives such as gradient diversity and gradient history consistency to identify high-value clients. Furthermore, we introduce a dynamic weighting mechanism that harmonizes the dual aggregation processes of cloud-edge and edge-client layers, enabling adaptive optimization for heterogeneous edge data distributions and achieving a better tradeoff between performance and fairness. Extensive experiments conducted under diverse heterogeneous scenarios and datasets demonstrate that the proposed method improves global model accuracy by an average of 12.9% and fairness by 26.6% compared with existing algorithms. Particularly, under strong heterogeneity conditions such as the CIFAR-10 dataset, G-M3CS + achieves a 23.3% performance improvement over baseline methods, effectively mitigating the training challenges induced by edge-level data heterogeneity.

Time series anomaly detection (TSAD) has long been a hot research topic in data mining due to its various applications. Recent studies challenge the effectiveness of popular deep learning methods for TSAD, suggesting their failure in detecting subtle and prolonged anomalies. Outlier Exposure (OE) and Masked Autoencoder (MAE) emerge as two promising paradigms (classification and reconstruction) for solving the above problems. However, OE-based methods are constrained by poor generalization, while MAE-based methods are limited by masking misalignment issues. To address these limitations, this paper proposes a novel framework, CoAD, which unifies the two paradigms to leverage their complementary strengths while mitigating their respective weaknesses. In this framework, the classification module generates probability-informed soft masks for the reconstruction module, which in turn alleviates the generalization problem of the classification module. This cooperative design enables CoAD to effectively detect subtle and complex anomalies that are often overlooked by existing methods. Additionally, the classification module is carefully designed to resolve issues related to improper classification granularity and the neglect of frequency information. Extensive experiments on high-quality benchmark datasets, conducted under rigorous evaluation protocols, demonstrate that CoAD significantly outperforms both state-of-the-art deep learning and traditional data mining methods, highlighting the potential of deep learning in TSAD. Moreover, CoAD is lightweight and substantially faster than existing SOTA methods, demonstrating its practical value for large-scale, real-time applications.

Temporal Knowledge Graph Question Answering (TKG-QA) requires reasoning over time-evolving facts, yet existing approaches struggle to jointly capture temporal dependencies and semantic understanding. Traditional graph-based methods often miss semantic nuances, while LLM-based solutions lack temporal alignment, leading to inaccurate retrieval and incomplete reasoning. To address this gap, we propose a framework that bridges structured temporal modeling with the reasoning capacity of large language models. Our approach reconstructs temporal knowledge into semantically coherent text representations and employs a hybrid retrieval mechanism to ensure precise and temporally consistent evidence selection. Besides, we design an interpretable reasoning process that guides LLMs to effectively combine temporal structure with semantic inference. Experiments on CronQuestions and MultiTQ demonstrate significant gains, with improvements of 6.1% and 12.1% over the best baselines, confirming the advantage of unifying temporal knowledge modeling and language-based reasoning for TKG-QA.

Time series anomaly detection remains one of the most active research areas in data mining due to its wide range of real-world applications. In recent years, numerous deep learning-based methods have been proposed for this task. However, deep learning-based methods fail to detect subsequence anomalies with long durations, lack explainability, and are vulnerable to training set contamination. This paper addresses these issues by proposing a novel deep learning framework for effective, explainable, and robust time series anomaly detection. Our framework, MMA, incorporates the MLP-Mixer backbone with a Masked Autoencoder-based anomaly detection approach to allow for a significantly larger input window size (10 to 20 times larger than the input window sizes of current models). This larger input window enables our model to detect challenging subsequence anomalies. Meanwhile, a contrast learning module is proposed to aid in detecting subtle anomalies that fail to be identified by residual errors. Furthermore, a dynamic anomaly filtering method is introduced to mitigate the impact of subsequence anomalies on the reconstruction of surrounding normal regions to reduce false alarms. Extensive experiments on univariate and multivariate time series datasets demonstrate that our proposed framework significantly outperforms state-of-the-art methods across rigorous evaluation metrics. Additionally, MMA has a strong ability to reconstruct potential normal patterns in anomalous regions, providing high levels of explainability. Moreover, MMA demonstrates high robustness to various types of training set pollution.

Text-to-image models can generate diverse and creative images under textual guidance, with broad applications in artistic creation, game design, image restoration, and other fields. However, current text-to-image models still exhibit shortcomings, particularly in scenarios requiring precise control over the positions of generated objects, where one or several specified subjects in the input prompt often fail to appear. To address these issues, this paper proposes an Iterative Optimization based on the Generation of Position Control Deficiencies (IOPCD) algorithm for position-controlled image generation. During the inference stage of the model, the IOPCD algorithm dynamically intervenes in the image generation process at each timestep by introducing iterative optimization for missing generation. This ensures that the model fully gives attention to all subjects in the input prompt during each iteration, thereby mitigating the issue of subject omission during inference and improving the semantic fidelity of the generated images. Comparative experiments on real-world datasets demonstrate that, compared to traditional position-controlled image generation algorithms, the proposed method achieves an average improvement of 4.7% in the mAP metric (measuring positional accuracy) and 8.2% in the FID metric (measuring generation omissions). The results confirm that the proposed method significantly outperforms traditional image generation models in precisely controlling the positions of subjects and alleviating subject omission issues.

Generalized Category Discovery (GCD) aims to leverage labeled old categories data and abundant unlabeled data to discover novel categories unseen during training. However, prior studies have revealed significant limitations in parametric classifiers for novel class recognition. This paper rethinks the role of self-labeling in parametric classification for GCD and identifies that while self-labeling introduces training biases, it holds untapped potential for boosting model performance. We conduct an in-depth analysis of the reason why self-labeling fails in parametric classification and propose a Decoupled Self-labeling (DSL) approach that separates pseudo-labels generation from the predictive classification head. This design minimizes the learning bias introduced by unreliable pseudo-labels while preserving their utility for feature learning. Additionally, we introduce a label encoding risk minimization strategy that provides supervised guidance for unlabeled further extract valuable information from pseudo-labels. Comprehensive evaluations on GCD benchmarks demonstrate that DSL achieves state-of-the-art performance, outperforming parametric and non-parametric baselines, especially in the novel categories.

Generative models have become an exciting area of research in recent years. Generative Adversarial Networks (GANs) and Denoising Diffusion Probabilistic Models (DDPMs) have been utilized in various data augmentation applications. However, generative models can learn high-dimensional features of data through adversarial learning, making them suitable for nonlinear and nonstationary time series analysis, such as stock market prediction, high-frequency trading, and ocean current forecasting. In this paper, the researchers focus on using a wavelet-based GAN to predict stock market prices by generating synthetic stock market price trends. Historical stock price data from 2014 to 2024 is used for our experiments, and the results show that the wavelet-based GAN outperforms deep learning baseline models..

In recent years, blockchain has emerged as a promising technology with extensive applications in various fields. One of its most notable applications is cryptocurrency. However, the prevalence of phishing scams in blockchain transaction networks has led to significant economic losses and poses a severe threat to transaction security within the cryptocurrency ecosystem. Existing methods for phishing scams detection often employ traditional machine learning techniques or graph embedding methods to extract key information that distinguishes phishing addresses. Nevertheless, these methods often overlook the temporal information within transaction networks, failing to fully capture the dynamic nature of the blockchain transaction network, resulting in suboptimal detection performance. In this paper, we propose a Temporal Graph Attention Network for blockchain phishing scams detection. Specifically, we use a Long Short-Term Memory (LSTM) network to obtain temporal transaction representations. Additionally, we utilize an attention mechanism to aggregate transaction features and features between neighboring nodes. Finally, by incorporating the obtained node representations and the topological characteristics of nodes, we identify phishing addresses using a Multilayer Perceptron (MLP). Experimental results on three real-world Ethereum phishing scams detection datasets indicate that our proposed method significantly outperforms competing approaches.

Federated Learning (FL) is a promising distributed machine learning framework that emphasizes privacy protection. However, inconsistencies between local optimization objectives and the global objective, commonly referred to as client drift, primarily arise due to non-independently and identically distributed (Non-IID) data, multiple local training steps, and partial client participation in training. The majority of current research tackling this challenge is mainly based on the empirical risk minimization (ERM) principle, while giving little consideration to the connection between the global loss landscape and generalization capability. This study proposes FedGAM, an innovative FL algorithm that incorporates Gradient Norm-Aware Minimization (GAM) to efficiently search for a local flat landscape. FedGAM specifically modifies the client model training objective to simultaneously minimize the loss value and first-order flatness, thereby seeking flat minima. To directly smooth the global flatness, we propose the more significant FedGAM-CV, which employs control variables to correct local updates, guiding each client to train models in a globally flat direction. Experiments on three datasets (CIFAR-10, MNIST, and FashionMNIST) demonstrate that our proposed algorithms outperform existing FL baselines, effectively finding flat minima and addressing the client drift problem.

Smart contracts are decentralized applications de-ployed extensively on blockchain. Due to their economic nature, vulnerabilities in smart contracts can lead to potential significant economic and property losses, disrupting the stable ecosystem of Ethereum. Therefore, the detection of smart contract vul-nerabilities is of paramount importance. Current mainstream methods for smart contract vulnerability detection rely on heuris-tic algorithms based on manual design, which lack reusability across different application scenarios, are time-consuming, and exhibit suboptimal accuracy. To enhance vulnerability detection effectiveness, a method tailored for timestamp vulnerabilities in smart contracts is proposed, named SESCD, based on self-ensembling pretraining. The proposed approach first identifies potential data propagation paths for timestamp vulnerabilities, prunes them, and leverages self-ensembling pretrained models to learn about these propagation paths. Furthermore, the training process is optimized through knowledge distillation to improve the model's ability to detect whether smart contracts contain timestamp vulnerabilities. SESCD demonstrates superior vulner-ability detection and generalization capabilities, alleviating performance instability issues caused by insufficient training data. To validate the effectiveness of SESCD, comparative experiments are conducted on a real-world dataset of smart contracts against 13 mainstream smart contract vulnerability detection methods. Experimental results show that SESCD achieves precision, recall, and F1 scores of 0.91, 0.93, and 0.92 respectively in detecting timestamp vulnerabilities. Compared to the 13 mainstream methods, SESCD exhibits an average relative improvement of 28%, 30%, and 30%, significantly enhancing the detection capabilities of timestamp vulnerabilities.

The rapid development of technologies such as the Internet of Things (IoT) has greatly facilitated various aspects of people's lives and profoundly transformed their daily routines and work environments. Sensor devices, a vital component of the IoT, offer a reliable and real-time data source for the advancement of applications like artificial intelligence and the industrial internet. However, there are still numerous unresolved issues in the data collection process of sensor devices, particularly in harsh conditions. Utilizing drone platforms for real-time data collection appears to be a viable approach at present. This paper addresses the optimization and control of energy consumption in unmanned aerial vehicle (UAV) data collection. It comprehensively considers factors such as the distribution of sensor devices, UAV speed control, and sensor energy consumption. By integrating Markov fluid theory, we provide a mathematical expression of the relationship between UAV energy consumption, the distribution of sensor devices, and data collection transmission variables in this scenario.

Entity resolution (ER), which aims to identify whether data records from various sources refer to the same real-world entity, is a crucial part of data integration systems. Traditional ER solutions assumes that data records are stored in relational tables with an aligned schema. However, in practical applications, it is common that data records to be matched may have different formats (e.g., relational, semi-structured, or textual types). In order to support ER for data records with varying formats, Generalized Entity Resolution has been proposed and has recently gained much attention. In this paper, we propose PromptER, a model based on pre-trained language models that offers an efficient and effective approach to accomplish Generalized Entity Resolution tasks. PromptER starts with a supervised contrastive learning process to train a Transformer encoder, which is afterward used for blocking and fine-tuned for matching. Specially, in the record embedding process, PromptER uses the proposed prompt embedding technique to better utilized the pre-trained language model layers and avoid embedding bias. Morever, we design a novel data augmentation method and an evaluation method to enhance the performance of the proposed model. We conduct experiments on the Generalized Entity Resolution dataset Machamp and the results show that PromptER significantly outperforms other state-of-art methods in the blocking and matching tasks.

Entity Matching (EM) aims to determine whether records in two datasets refer to the same real-world entity. Existing work often uses Pre-trained Language Models (PLMs) for feature representation, converting EM to a binary classification task. However, due to the dependence of PLMs on large labeled datasets and the overlap between train and test sets in current EM benchmarks, these methods often underperform in real-world scenarios (e.g., small data size, hard negative samples, and unseen entities). To address the limitations of existing methods, we propose SETEM, a self-ensemble training method leveraging the stability and strong generalization of ensemble models to tackle these challenges in real-world scenarios. Additionally, we develop a faster training method for low-resource applications. Experiments on benchmark datasets show that SETEM outperforms Ditto and HierGAT on the F1 score. In particular, SETEM shows the greatest advantage with small datasets and a high proportion of unseen entities in the test set, achieving up to a 9.61% F1 score increment over baselines on the WDC dataset.

In the real world, collecting time series data often encounters difficulties and uncertainties, leading to prevalent data missing issues. The time series imputation task aims to restore data integrity and improve prediction performance. Existing models mainly use single imputation to handle missing data. This paper proposes a Patch-based Multiple Imputation Transformer (PMIT), which learns data correlations within the same feature through patch operations and channel-independent (CI) operations. The model employs multiple imputation training, optimizing imputed data propagation across different modules, and incorporates attention mechanisms and dynamic adjustment methods to address complex dependencies. Experimental results show that PMIT outperforms recent models in imputation performance and generalization ability across various realworld datasets.

面向大规模物联网系统高动态、运行环境复杂等特征,构建了基于机会网络的存储-携带-转发信息传输模式,进而提出了带有传输不确定性的信息传输性能评估模型.在此基础上,综合考虑信息传输过程的能量消耗与传输性能等目标,提出了基于庞特李雅金极大值定理的最优传输策略.进一步,通过理论分析证明了最优传输策略服从阈值形式.仿真实验表明,模型的误差低于4.0 8％.同时,实验也表明所提出的最优传输策略明显优于传统的静态策略,并随着信息有效期的递增,性能表现更好.

Question answering (QA) systems are designed to answer questions based on given information or with the help of external information. Recent advances in QA systems are overwhelmingly contributed by deep learning techniques, which have been employed in a wide range of fields such as finance, sports and biomedicine. For generative QA in open-domain QA, although deep learning can leverage massive data to learn meaningful feature representations and generate free text as answers, there are still problems to limit the length and content of answers. To alleviate this problem, we focus on the variant YNQA of generative QA and propose a model CasATT (cascade prompt learning framework with the sentence-level attention mechanism). In the CasATT, we excavate text semantic information from document level to sentence level and mine evidence accurately from large-scale documents by retrieval and ranking, and answer questions with ranked candidates by discriminative question answering. Our experiments on several datasets demonstrate the superior performance of the CasATT over state-of-the-art baselines, whose accuracy score can achieve 93.1% on IR&QA Competition dataset and 90.5% on BoolQ dataset.

Fingerprint is one of the most widely used biometric features to identify a person. Fingerprint matching is to compare an input fingerprint with fingerprints within the database to find the most similar fingerprints. However, existing fingerprint matching technologies face two major challenges: data quantity and data quality. From the perspective of data quantity, the existing fingerprint database stores hundreds of millions of data, so the traditional iterative matching method is not practical due to huge time consumption. From the data quality point, existing fingerprint products have a small collection area for aesthetics and portability, resulting in a low quality of the captured finger-print image, which seriously affects the accuracy of the matching algorithms. To address the above problems, we present a two-stage fingerprint matching method, which not only improves the efficiency of fingerprint matching, but also ensures the accuracy of fingerprint matching. The first stage utilizes a convolutional neural network trained with triplet loss to extract the overall feature of a fingerprint image. The cosine similarity between features can be used to identify possible fingerprint matches. The second stage compares fingerprint matches acquired in the first stage using a minutiae points comparing algorithm to get a more precise result. The experiment result shows that our method is satisfactory in both speed and accuracy.