School bullying represents a critical societal challenge with lasting psychological and academic consequences for affected students. Despite recent advances in machine learning for predicting bullying behaviors, conventional models struggle to capture the complex, non-linear relationships among contributing factors, especially under the imbalanced data distributions typical of real-world bullying cases. Furthermore, the inherent opacity of Deep Neural Networks (DNNs) restricts their application in educational contexts where interpretability and actionable insights are essential. In this paper, we propose a novel automated framework that integrates Neural Architecture Search (NAS) with Shapley value-based explanation methods to jointly address performance and interpretability challenges. Our framework automatically identifies optimal DNN architectures tailored for bullying prediction, incorporating mechanisms to handle class imbalance without extensive manual tuning. To address model transparency, we employ a Shapley value analysis pipeline that systematically attributes predictions to key risk factors, offering educators and policymakers principled and quantitative insights. Extensive experiments on publicly available datasets demonstrate that our method significantly outperforms state-of-the-art baselines, achieving notable improvements in Accuracy (+2.58 %), F1-Score (+34.52 %), and AUC (+6.47 %). Importantly, the feature importance rankings from our Shapley analysis closely align with established sociological and educational theories on bullying, affirming the model's interpretability and practical relevance. Cross-dataset validation further verifies the framework's generalizability to broader youth behavioral risk prediction tasks. Our code is submitted at https://github.com/clsyc/Bullyingshapley.

Federated learning, as a distributed machine learning paradigm, is prevalent in various fields but suffers from security issues. On the server side, unverified aggregation allows the attacker to manipulate model parameters, while on the client side, the malicious clients can deliver the poisoning gradients, both rendering the aggregation useless. As a response, current studies bring in verifiable mechanisms to ensure the integrity of server-side aggregation, whereas illegitimate clients can still participate training in client-side without authentication. In this paper, we propose a Bidirectional Verifiable Secure Aggregation Protocol (BVSAP), which enhances the security of federated learning with theoretical guarantees. BVSAP ensures both the correctness of the aggregation results on the server and the legitimacy of participating clients through verifiable secret sharing. In BVSAP, homomorphic encryption serves to protect the entire aggregation process from potential model disclosure attacks. By avoiding the use of highly complex operations such as bilinear pairing in the verification process, BVSAP reduces the computational complexity of verification. The experimental results across MNIST and CIFAR-100 demonstrate the soundness of our methods.

Cardiovascular disease represents the leading cause of mortality in China, accounting for over 40% of all deaths. Existing risk prediction models predominantly derive from Western populations, rendering them suboptimally calibrated for the Chinese middle-aged and elderly demographic. Conventional statistical approaches inadequately capture non-linear associations within high-dimensional data, whilst machine learning models, despite superior performance, suffer from insufficient interpretability. This study leveraged a nationally representative cohort to develop an interpretable machine learning-based tool for long-term cardiovascular risk prediction tailored to the Chinese population. To compare the predictive performance of ten machine learning algorithms using data from the China Health and Retirement Longitudinal Study (CHARLS), identify the optimal model, achieve transparent interpretation through SHapley Additive exPlanations (SHAP) methodology, and develop an individualized cardiovascular risk assessment tool for Chinese residents aged 45 years and above. The study enrolled 8,080 participants aged ≥ 45 years without baseline cardiovascular disease from the CHARLS 2011-2020 longitudinal dataset, with 9-year follow-up. The primary outcome was incident cardiovascular disease. From 77 candidate variables, logistic regression analysis identified 11 predictors: geographical region, hypertension, dyslipidaemia, liver disease, asthma, depression score, age, sleep duration, triglycerides, high-density lipoprotein cholesterol, and waist circumference. The cohort was randomly partitioned into training (n = 5,657, 70%) and validation (n = 2,423, 30%) sets. Ten predictive models were constructed, including random forest, gradient boosting machine, and extreme gradient boosting. Model performance was evaluated using area under the receiver operating characteristic curve (AUROC), calibration plots, and decision curve analysis. Feature contributions were elucidated using SHAP values. Incident cardiovascular disease occurred in 1,246 participants (22.0%) within the training cohort. Multivariable analysis identified hypertension (adjusted OR 1.80), waist circumference (adjusted OR 1.05 per 1-cm increment), dyslipidaemia (adjusted OR 1.42), and liver disease (adjusted OR 1.60) as principal independent predictors. Among ten algorithms evaluated, random forest demonstrated superior performance: validation set AUC 0.829 (95% CI 0.809-0.848), accuracy 0.770, sensitivity 0.681, specificity 0.795. The model exhibited excellent calibration and yielded maximal net clinical benefit across the 10%-85% risk threshold spectrum. SHAP analysis revealed waist circumference as the predominant contributor, followed by triglycerides, age, and hypertension. Psychobehavioural factors (depression, sleep duration) demonstrated independent predictive value. A web-based risk calculator was developed, providing real-time individual 9-year cardiovascular disease probability estimates. The random forest model accurately predicts cardiovascular disease risk in the Chinese middle-aged and elderly population, with waist circumference emerging as the most critical predictor. Translated into an online assessment tool, this model facilitates community-based screening and individualized prevention, offering a pragmatic risk stratification approach for resource-constrained settings.

It is difficult for clustering to measure the similarity or dissimilarity between two categorical data objects because the categorical data lack a clear space structure. Some researchers proposed methods to map categorical objects into Euclidean space to enhance the distinguishability of objects. However, the existing spatial mapping methods and related categorical data clustering algorithms only consider the object level and do not exploit the attributes of the categorical data to reduce the datasets. And when the dataset is large, the current algorithms are very time-consuming. Categorical data attributes inherently partition dataset. Therefore, it is necessary to analyse data distributions at the attribute level in order to select reference sets that more appropriately represent the distribution of categorical data and to construct the space structure of categorical data. In this paper, a new clustering algorithm for categorical data with Mean Gain Ratio (MGR) based reference set selection method is proposed. In detail, firstly, a MGR based method is given for selecting a more appropriate reference set. This method first selects the attribute with the highest MGR, and then selects an object from each equivalent class of the partition generated by that attribute to form a reference set. And then, we present a clustering algorithm for categorical data by combining the proposed MGR based representation and the k-means algorithm. The results of comparative experiments show that the proposed method achieves higher clustering performance than existing methods. Furthermore, when dealing with very large datasets, the algorithm proposed in this paper performs better in terms of time complexity and scalability.

We present DuPO, a dual learning-based preference optimization framework that generates annotation-free feedback via the generalized duality. DuPO addresses two key limitations: Reinforcement Learning with Verifiable Rewards (RLVR)’s reliance on costly labels and applicability restricted to verifiable tasks, and traditional dual learning’s restriction to strictly dual task pairs (e.g., translation and back-translation). Specifically, DuPO decomposes a primal task’s input into known and unknown components, then constructs its dual task to reconstruct the unknown part using the primal output and known information (e.g., reversing math solutions to recover hidden variables), broadening applicability to non-invertible tasks. The quality of this reconstruction serves as a self-supervised reward to optimize the primal task, synergizing with LLMs’ ability to instantiate both tasks via a single model. Empirically, DuPO achieves substantial gains across diverse tasks: it enhances the average translation quality by 2.1 COMET over 756 directions, boosts the mathematical reasoning accuracy by an average of 6.2 points on three challenge benchmarks, and enhances performance by 9.3 points as an inference-time reranker (trading computation for accuracy). These results position DuPO as a scalable, general, and annotation-free paradigm for LLM optimization.

In this article, we propose a random projection-based response best-subset selector to perform response variable selection in ultra-high dimensional multivariate data, where both the dimensions of response and predictor variables are substantially greater than the sample size. This method is developed by integrating the response best-subset selector and random projection technique which is applied to reduce dimensionality of predictors. Under a multivariate tail eigenvalue condition, such a random projection-based dimensionality reduction of predictors only leads to an ignorable error between the original and dimension-reduced models. A computational procedure is presented. The proposed method exhibits model consistency under some certain conditions. The efficiency and merit of the proposed method are strongly supported by extensive finite-sample simulation studies. A real breast cancer dataset spanning 22 chromosomes are analyzed to demonstrate the proposed method.

To fully utilize the advantage of reservoir computing in deep network modeling, a deep echo state network with projection-encoding (DEESN) is newly proposed for multi-step time series prediction in this paper. DEESN integrates multiple echo state network (ESN) modules and extreme learning machine (ELM) encoders in series arrays. Firstly, the kth ESN in DEESN learner is responsible for kth step ahead prediction. The forecast output and encoded reservoir states of the previous ESN module are concatenated with the input variable to form the new input signals for the next adjacent module. Therefore, the temporal dependency among future time steps can be learned, which contributes the performance improvement. Secondly, the ELM encoder is used to optimize the reservoir states for time consumption reduction. Finally, the effectiveness of DEESN is evaluated in artificial chaos benchmarks and real-world applications. Experimental results on six different datasets and comparative models demonstrate that the proposed DEESN has excellent accuracy and robust generalization for multi-step time series prediction.

Existing fashion recommendations often rely on natural language processing or content-based image retrieval, overlooking direct aesthetic assessments of fashion images. Given the subjectivity and complexity of this task, we propose treating fashionability as a relative attribute to rank paired clothing images. To address this ranking challenge, we propose Ordinal Focal Loss, which transforms the pairwise ranking problem into a multi-classification task, leveraging ordinal attributes to improve classification boundaries. Furthermore, in terms of fashion feature representation, we propose modeling not just individual items but also their combined effect as an outfit, providing a more holistic and nuanced fashion representation. We introduce the Fashionability3k dataset, comprising 3k image pairs (2398 ordered and 601 similar pairs) with objective relative fashion labels. Experiments on three datasets—our Fashionability3k and two public datasets—show that our method outperforms the baseline by nearly 1 % in ranking accuracy. In the user study, it achieved a 0.72 consistency with human subjective perception. Moreover, combining local and global visual features leads to additional performance gains, with an average improvement of 2.78 % in ordered pairs and 1.09 % in similar pairs. This is the first study to treat fashionability as an objective attribute for comparative analysis, validated through extensive experiments.

Multilingual translation stands as a challenging task for large language models (LLMs) to handle intricate language patterns and stilted translations that arise in automated translations. In this paper, we introduce Seed-X, a family of open-source LLMs comprising instruct and reasoning models, pushing the limits of translation capability with 7B parameter size. The base model is pre-trained on a diverse, high-quality dataset encompassing both monolingual and bilingual content across 28 languages, harnessing the full potential of multilingual data. The instruct model is then finetuned to translate by Chain-of-Thought (CoT) reasoning and further enhanced through reinforcement learning (RL) to achieve better generalization across diverse language pairs. Seed-X achieves performance comparable to leading closed-source models, including Gemini-2.5 and GPT-4o, across 28 languages, and significantly outperforms larger open-source models in both automatic metrics and human evaluations. We share the best practices through our optimization process, and make the parameter public available for advancing translation research and applications.

In order to ensure the stable transmission of electric power, it is an effective way to diagnose and maintain the operating status of electric power equipment from the operation data of electric power equipment. This paper uses a stacked sparse autoencoder to design a training model to realize the data function operation function in the fault detection model. After collecting and classifying the power system data, the line current is standardized and transformed. Then, the processed data is input into the stacked sparse autoencoder, and the model is trained layer by layer. On this basis, the long-term memory network model is introduced to establish a fault diagnosis model. To solve the double-sample situation of power data, the maximum mean difference method must be used. A preventive maintenance strategy is constructed based on failure prediction and remaining life to optimize the implementation path. Evaluate the model’s value in terms of its performance, reliability, and economic benefits of preventive O&M methods. However, judging from the fuzzy fault degree, the electrical components with a high probability of failure are T1, T3, L3, L9, and B1, and the fuzzy fault degrees are 0.3154, 0.2789, 0.0648, 0.2657, and 0.0678, respectively—fusion of multidimensional evidence. The components most likely to fail are T1, T3, and L9. From the perspective of operation and maintenance costs, when the MMC maintenance time of Dublin Fang Electric Farm is 17 times, the lowest operation and maintenance cost is 22.8615 million yuan.

Physical education is an important way to strengthen the physical quality of students in the university. Sport performance is an important criterion for judging the physical fitness of college students. But during the process of calculating and collecting the comprehensive score for each student, teachers will use different equipment or various grading formulas to evaluate the grade for many years. Thus, referenced score will lose their values if the measurement data for each student in different years are not unified and sometimes with the subjective factor because of the manual calculations. This article analyzes the details of physical fitness test results and the relationship with the comprehensive score. It is discussed that the comprehensive score can provide serious help for teachers to know students’ physical quality level and make reasonable teaching program to each student according to their personal radar chat. Therefore, the artificial influence is excluded, and the prediction model is designed by using the principal component analysis method and the back propagation (BP) neural network technology. The predictive model allows students to evaluate their own physical test score in advance, get a preliminary understanding of their physical fitness. Comparisons are made between the applications of this model in different years and errors are discussed to verify the accuracy. The results indicate that the comprehensive score prediction model supplies one effective approach to unify the scoring standards and improve the computation efficiency in physical education.

Load forecasting (LF) models are essential for various smart grid applications, and their accuracy heavily relies on the quality of input load data and load types. Previous LF studies have ignored noise loads due to tampering, transmission failures, etc., and have not considered the fusion of different types of loads, both of which have an impact on load forecasting accuracy. To address these issues, this study introduces a novel multi-type load forecasting model named MLF-RFDC, based on random forest and density clustering, that enjoys three-fold ideas: (1) it treats load data from each electrical activity as an independent data matrix, capturing variation patterns unique to each load type; (2) it identifies and corrects noisy entries in each data matrix using a low-rank clustering approach, highlighting noises as outliers and restoring them through latent factor analysis; and (3) it combines noise-free data matrices from all load types to enhance LF accuracy from an ensemble perspective. Extensive experiments conducted on ten benchmark datasets and three real-world load datasets demonstrate that our proposed algorithm outperforms 11 state-of-the-art models. Specifically, the performance results are remarkable: (1) the anomaly data detection accuracy is enhanced by up to 15.66%; (2) the Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) for anomaly data recovery show significant improvements of tens of times; and (3) the MAE, RMSE, MAPE, and R-squared (R2) for load forecasting are the most favorable.

In order to understand the multidimensional situation prediction of digital twin active power grid, a research on multidimensional situation prediction of digital twin active power grid based on LSTM algorithm is proposed. In this paper, firstly, a multi-dimensional situation prediction algorithm of power grid key indicators based on LSTM is established to realize the change prediction of key indicators attributes of digital twin active power grid. Secondly, the data of several key indicators such as load characteristics are collected, and a multi-dimensional system prediction model is established, which can control the state of active power grid; The LSTM prediction algorithm is proposed to fit the characteristics of multi-dimensional data, and the next stage of multi-dimensional data prediction is mapped to the power digital twin, so as to realize the synchronous implementation and intelligent regulation of smart energy system operation planning. Finally, a simulation test model is established, and an example shows that the multi-dimensional situation prediction method of digital twin power grid based on deep learning can better predict and distinguish the power grid situation, and provide decision support for accurate planning of energy system in the future.

As one of the most popular software applications, a web application is a program accessible through the web that dynamically generates content based on user interactions or contextual data; examples include online shopping platforms, social networking sites, and financial services. Web applications operate in diverse environments and leverage web technologies such as HTML, CSS, JavaScript, and Ajax, often incorporating features like asynchronous operations to enhance user experience. Due to the growing number of users and the popularity of web applications, the quality of these applications has become increasingly important. Web Application Testing (WAT) plays a vital role in ensuring the functionality, security, and reliability of web applications. Given the speed with which web technologies are evolving, WAT is especially important. In the last twelve years, various WAT approaches have been developed. The diversity of approaches reflects the many aspects of web applications, such as dynamic content, asynchronous operations, and diverse user environments. This paper provides a comprehensive overview of the main achievements over the last twelve years: It examines the main steps involved in WAT, including test case generation and execution, as well as evaluation and assessment. The currently available tools for WAT are also examined. The paper also discusses open research challenges and potential future work in WAT.

Because a considerable number of college graduates employment situation is bad, and in trouble, so the decision tree algorithm research to help deal with. In this study, the concept of decision tree algorithm is introduced, and then the selection of algorithm is carried out. Based on the algorithm of research institute selection, the employment data of graduates is analyzed. Through the research, the causes of the employment problems of college graduates can be understood by the relevant decision tree algorithm, and the employment direction can be given to improve the employment situation.

Because of the problems that the fast k-Medoids clustering algorithm does not consider the weight of each attribute and the initial clustering center may be in the same cluster, this paper proposes a weighted $\boldsymbol{k}$-Medoids clustering algorithm based on granular computing. Firstly, the hierarchical structure in the fuzzy quotient space theory is introduced to define the decision attribute of the sample under each granularity, and the computing method of sample attribute weight is defined by the attributes of the sample set itself and the definition of attribute importance in the rough set model. Secondly, the sample similarity function is defined by the attribute weight coefficient, and the attribute weight is integrated into the similarity of the fast k-Medoids clustering algorithm to quantitatively define the importance of each sample's attribute. Finally, from the prospective view of granular computing, the samples are clustered according to the above similarity function, and the original clustering centers are initialized by K cluster centers with long distance. The experimental results on machine learning datasets UCI show that the proposed weighted k-Medoids clustering algorithm based on granular computing greatly improves the accuracy of clustering.

Prompt tuning attempts to update few task-specific parameters in pre-trained models. It has achieved comparable performance to fine-tuning of the full parameter set on both language understanding and generation tasks. In this work, we study the problem of prompt tuning for neural text retrievers. We introduce parameter-efficient prompt tuning for text retrieval across in-domain, cross-domain, and cross-topic settings. Through an extensive analysis, we show that the strategy can mitigate the two issues -- parameter-inefficiency and weak generalizability -- faced by fine-tuning based retrieval methods. Notably, it can significantly improve the out-of-domain zero-shot generalization of the retrieval models. By updating only 0.1% of the model parameters, the prompt tuning strategy can help retrieval models achieve better generalization performance than traditional methods in which all parameters are updated. Finally, to facilitate research on retrievers' cross-topic generalizability, we curate and release an academic retrieval dataset with 18K query-results pairs in 87 topics, making it the largest topic-specific one to date.

In recent years, thanks to the development of GPU (Graphic Processing Unit) computing power, deep neural network (DNN) has drawn extensive attention from academia and industry and has been widely used in various fields of artificial intelligence. However, the mainstream deep learning models have some problems such as large storage space consumption and high computational complexity, which makes it difficult for the deep neural network model to be deployed to mobile devices with limited computing resources or applications with strict delay requirements. Therefore, on the premise of minimizing the loss of model accuracy, the algorithm researches of deep neural network compression and acceleration which can lighten the model has been carried out. In recent years, model compression methods such as pruning, quantification and knowledge distillation have emerged. This paper focuses on neural network pruning algorithm. Firstly, this paper expounds the basic methods and classification of neural network pruning algorithms. Then according to the classification of pruning granularity, a more detailed classification, analysis and comparison of specific models are carried out from the algorithm level. Then based on cifar-10 data set, validity tests of some neural network pruning algorithms and contrastive analyses of the experimental data from different models such as the parameters are carried out. Finally, in combination of the recent results of neural network pruning algorithm, we put forward research dilemmas and problems to be solved. Then the future research trends and methods of neural network pruning algorithm are prospected.