Motivated by the analysis of data from a clinical trial on patients with early breast cancer, we propose in this paper a new joint model that uses a Tobit partly linear mixed model for longitudinal measurements which are bounded in an interval and have a nonlinear relationship with the observation times and a semiparametric mixture cure model that incorporates a B-spline baseline hazard for survival times with cure proportion. A procedure is developed for estimating parameters in the proposed model using the partial likelihood and Laplace approximation. Additionally, a method of random weighting is proposed to compute the variances of the parameter estimators. The performance of the proposed model and the inference procedures is evaluated through simulation studies and data from the clinical trial that motivated this study.

Currently, many researchers use weights to merge self-matched words obtained through dictionary matching in order to enhance the performance of Named Entity Recognition (NER). However, these studies overlook the relationship between words and sentences when calculating lexical weights, resulting in fused word information that often does not align with the intended meaning of the sentence. Addressing above issue and enhance the prediction performance, we propose an adaptive lexical weight approach for determining lexical weights. Given a sentence, we utilize an enhanced global attention mechanism to compute the correlation between self-matching words and sentences, thereby focusing attention on crucial words while disregarding unreliable portions. Experimental results demonstrate that our proposed model outperforms existing state-of-the-art methods for Chinese NER of MRSA, Weibo, and Resume datasets.

This article proposes a hybrid model to assist insurance companies accurately assess the risk of increasing claims for their premiums. The model integrates long short-term memory (LSTM) networks and convolutional neural networks (CNN) to analyze historical claim data and identify emerging risk trends. We analyzed data obtained from insurance companies and found that the hybrid CNN-LSTM model outperforms standalone models in accurately assessing and categorizing risk levels. The proposed CNN-LSTM model achieved an accuracy of 98.5%, outperforming the standalone CNN (95.8%) and LSTM (92.6%). We implemented 10-fold cross-validation to ensure robustness, confirming consistent performance across different data splits. Furthermore, we validated the model on an external dataset to assess its generalizability. The results demonstrate that the model effectively classifies insurance risks in different market environments, highlighting its potential for real-world applications. Our study contributes to the insurance industry by providing valuable insights for effective risk management strategies and highlights the model’s broader applicability in global insurance markets.

Structured pruning is a standard tool for compressing deep neural networks, but its practical performance depends on how sparsity is allocated across layers. We propose FAIR-Pruner, a search-free framework for adaptive layer-wise structured pruning. FAIR-Pruner uses two within-layer rankings: a removal-oriented signal that proposes candidate units and a protection-oriented signal that identifies task-sensitive units. Its core component, Tolerance of Difference (ToD), measures the overlap between the removal prefix and the protected tail, and uses a shared tolerance level to induce non-uniform pruning depths across layers. As a default vision instantiation, FAIR-Pruner combines a Wasserstein-based U-Score for class-conditional unit separability with a Taylor-based R-Score for task-level sensitivity; the same ToD allocation rule can also be paired with alternative removal signals. Theoretically, we analyze ToD through the population R-Score, derive rank-based control of the high-R-Score mass entering the pruning set, and identify an additive exchange condition for same-budget comparison with uniform pruning. Experiments on CIFAR-10, CIFAR-100, SVHN, and ImageNet across VGG, ResNet, DenseNet, ConvNeXt, and DeiT show strong accuracy–compression trade-offs. Prune-only experiments on routed-expert Qwen1.5-MoE-A2.7B-Chat further examine architectural extensibility under matched expert budgets. FAIR-Pruner is released as a pip-installable open-source package.

This paper proposes a hybrid Convolutional Neural Network (CNN) and Linear Discriminant Analysis (LDA) model that combines a deep convolutional neural network and linear discriminant analysis to investigate anti-selection risk in insurance markets. The model enhances risk assessments using extensive data from insurance companies' big data sources and advanced machine learning algorithms. This improves the detection of anti-selection tendencies and enhances overall risk management techniques. After the final convolution layer, we add a Linear Discriminant Analysis layer to the backbone model Convolutional Neural Network. The Linear Discriminant Analysis layer allows the model to gather features, minimizing variation within each class and maximizing separation between different classes. After the Linear Discriminant Analysis layer, we append a fresh, fully connected (FC) layer with softmax activation and made comprehensive adjustments. We employ both Convolutional Neural Network and Linear Discriminant Analysis models to extract features and perform classification. The hybrid Convolutional Neural Network (CNN) and Linear Discriminant Analysis (LDA) model demonstrate superior reliability, with a test accuracy score of 97.4%, surpassing the classification accuracy of the Convolutional Neural Network and Linear Discriminant Analysis models with 90.2% and 91.3%, respectively.

This paper considers a risk model driven by a spectrally negative L & eacute;vy process, where any surplus above b (0 < b < infinity) is deducted away as dividends and any deficit is covered by injected capitals/raised money. For such a risk model, we define a variant of Parisian ruin time as the first time that the surplus process stays continuously below a (0 < a < b < infinity) for a time interval with length larger than some pre-specified exponential random variable that is marked on this time interval. A recursive formula for the moments of the Net Present Value (NPV) of dividends paid until Parisian ruin is provided. The expected NPV of capitals injected until the Parisian ruin time is also characterized compactly in terms of the scale functions of the underlying process.

Multivariate control charts are commonly used in manufacturing engineering to identify abnormal changes in the process. In a high-dimensional paradigm, where the number of variables (p) exceeds the number of Phase I subgroups (m), the sample covariance matrix is ill-conditioned and will become singular. This situation makes the classical T-2 control chart inefficient and even invalid to employ for monitoring a mean vector in statistical process control. This study proposes a new multivariate shrinkage-based diagonal T-2 control chart for high dimensional data, where p is very large compared to m with individual observation. A shrinkage estimator is used to estimate a diagonal covariance matrix to obtain an invertible, well-conditioned, and efficient estimate of the sample covariance matrix. The proposal's performance is also evaluated using the probability of detecting a shift. A simulation study reveals that the proposed control chart has an efficient sensitivity in detecting shifts for the high dimensional data. Results also show that m has a minor effect on the probability of shift detection for out-of-control data. In general, this probability improves for larger p, when shift is introduced in all p.

Addressing the pressing challenge of insurance fraud, which significantly impacts financial losses and trust within the insurance industry, this study introduces an innovative automated detection system utilizing ensemble machine learning (EML) algorithms. The approach encompasses four strategic phases: 1) Tackling data imbalance through diverse re-sampling methods (Over-sampling, Under-sampling, and Hybrid); 2) Optimizing feature selection (Filtering, Wrapping, and Embedding) to enhance model accuracy; 3) employing binary classification techniques (Bagging and Boosting) for effective fraud identification; and 4) applying explanatory model analysis (Shapley Additive Explanations, Break-down plot, and variable-importance Measure) to evaluate the influence of individual features on model performance. Our comprehensive analysis reveals that while not every re-sampling technique improves model performance, all feature selection methods markedly bolster predictive accuracy. Notably, the combination of the Gradient Boosting Machine (GBM) algorithm with NCR re-sampling and GBMVI feature selection emerges as the most effective configuration, offering superior fraud detection capabilities. This study not only advances the theoretical framework for combating insurance fraud through AI but also provides a practical blueprint for insurance companies aiming to incorporate advanced AI strategies into their fraud detection arsenals, thereby mitigating financial risks and fostering trust systems.

Agricultural and livestock farming are important contributors to greenhouse gas (GHG) emissions, with methane emissions from ruminants being particularly significant. This research aims to showcase the innovative development of a digital twin platform integrated with AI for analyzing both current and historical GHG emissions. The digital twin harnesses AI for advanced predictive analytics, enabling the tracking and understanding of broad GHG emission trends over time. Key features of the platform include AI and machine learning models tailored for informative GHG estimation. It also offers interactive maps for visualizing spatial data and conducting association analysis. Users can benefit from dynamic historical data comparisons and utilize livestock emission calculators aligned with Intergovernmental Panel on Climate Change (IPCC) guidelines. The project supports Net Zero efforts by empowering farmers and organizational stakeholders to enhance environmental stew-ardship and promote sustainable agriculture.

In this paper, a family of geometric shrinkage variances estimators is derived under a new framework of Riemann manifold. Two approaches are proposed to estimate the optimal shrinkage parameter that minimizes a loss function defined as the square of a Riemann metric. Simulation studies show that the proposed estimators with the estimated optimal shrinkage parameter perform well under the percentage related improvement in average loss (PRIAL) criterion, which measures the degree of improvement over the naive sample variance estimators. Finally, the proposed estimators are applied to analyse a real data set of gene expression levels of cancer patients.

Fraudulent activities especially in auto insurance and credit card transactions impose significant financial losses on businesses and individuals. To overcome this issue, we propose a novel approach for fraud detection, combining convolutional neural networks (CNNs) with support vector machine (SVM), k nearest neighbor (KNN), naive Bayes (NB), and decision tree (DT) algorithms. The core of this methodology lies in utilizing the deep features extracted from the CNNs as inputs to various machine learning models, thus significantly contributing to the enhancement of fraud detection accuracy and efficiency. Our results demonstrate superior performance compared to previous studies, highlighting our model’s potential for widespread adoption in combating fraudulent activities.

Knowing expected milk yield can help dairy farmers in better decision-making and management. The objective of this study was to build and compare predictive models to forecast daily milk yield over a long duration. A machine-learning pipeline was provided and five baseline models as well as a novel stacking model were developed for the prediction of milk yield on the CowNflow dataset using 414 Holstein cattle records collected from 1983 to 2019. Four different feature selection methods were performed to evaluate the essential biological characteristics and feeding-related features which affect milk yield. The results showed that the overall performance of predictive models improved after proper feature selection, with an R 2 value increased to 0.811, and a root mean squared error (RMSE) decreased to 3.627. The stacking model achieved the best performance with an R 2 value of 0.85, a mean absolute error (MAE) of 2.537 and an RMSE of 3.236. This research provides benchmark information for the prediction of milk yield on the CowNflow dataset and identifies useful factors such as dry matter (DM) intake and lactation month in long-term milk yield prediction.

Multiple resolutions arise across a range of explanatory features due to domain-specific structures, leading to the formation of feature groups. It follows that the simultaneous detection of significant features and groups aimed at a specific response with false discovery rate (FDR) control stands as a crucial issue, such as the spatial genome-wide association studies. Nevertheless, existing methods such as the multilayer knockoff filter (MKF) generally require a uniform detection approach across resolutions to achieve multilayer FDR control, which can be not powerful or even not applicable in several settings. To fix this issue effectively, this article develops a novel method of stabilized flexible e-filter procedure (SFEFP), by constructing unified generalized e-values, developing a generalized e-filter, and adopting a stabilization treatment. This method flexibly incorporates a wide variety of base detection procedures that operate effectively across different resolutions to provide stable and consistent results, while controlling the false discovery rate at multiple resolutions simultaneously. Furthermore, we investigate the statistical theories of the SFEFP, encompassing multilayer FDR control and stability guarantee. We develop several examples for SFEFP such as eDS-filter and eDS+gKF-filter. Simulation studies demonstrate that the eDS-filter effectively controls FDR at multiple resolutions while either maintaining or enhancing power compared to MKF. The superiority of the eDS-filter is also demonstrated through the analysis of HIV mutation data.

This paper proposes a new adaptive approach called adaptive mean (ADM) that combines the strengths of trimming and winsorization to minimize the mean square error (MSE). ADM will contribute significantly to insurers by enabling them to mitigate risk exposure caused by inaccurate premium estimation and establish more precise premiums for their clients. Furthermore, the proposed mean offers several pros compared to the conventional mean. It facilitates the natural and intuitive calculation of risk loading, identifies and measures the impact of large claims, and analyzes premium susceptibility to skewed risk by capturing the tails of the relevant loss models. In addition, this paper reviews two established, robust credibility methods (trimming and winsorization). It compares them with our proposed method, resulting in a more reliable and robustly accurate method for credibility premium estimation. We presented and analyzed two real data sets from engineering insurance companies and the Wisconsin Office of the Insurance Commissioner to highlight the advances of ADM in minimizing the MSE and building more robust credibility models that are less vulnerable to outlier events and model uncertainty.

Motivated by recent advances made in the study of dividend control and risk management problems involving the U.S. bankruptcy code, in this paper we follow [44] to revisit the De Finetti dividend control problem under the reorganization process and the regulator’s intervention documented in U.S. Chapter 11 bankruptcy. We do this by further accommodating the fixed transaction costs on dividends to imitate the real-world procedure of dividend payments. Incorporating the fixed transaction costs transforms the targeting optimal dividend problem into an impulse control problem rather than a singular control problem, and hence computations and proofs that are distinct from [44] are needed. To account for the financial stress that is due to the more subtle concept of Chapter 11 bankruptcy, the surplus process after dividends is driven by a piece-wise spectrally negative Lévy process with endogenous regime switching. Some explicit expressions of the expected net present values under a double barrier dividend strategy, new to the literature, are established in terms of scale functions. With the help of these expressions, we are able to characterize the optimal strategy among the set of admissible double barrier dividend strategies. When the tail of the Lévy measure is log-convex, this optimal double barrier dividend strategy is then verified as the optimal dividend strategy, solving our optimal impulse control problem.

Infrared small target detection is widely applied in military and civilian fields. Due to the small size of infrared targets, textural detail is missing. Common target detection methods extract semantic feature by narrowing down the feature map several times, which may lead to the small targets lost in deep layers and are not effective for infrared small target detection. To solve this problem, we propose a novel network called deep asymmetric extraction and aggregation. The network mainly consists of two processes - the vertical feature extraction and the horizontal feature aggregation, both of which are enhanced by an asymmetric attention mechanism. In the vertical process, the use of asymmetric attention mechanism combined with the reduction of down-sampling makes the small target better retained in the deep layers. Then through the horizontal process, shallow spatial feature and deep semantic feature are aggregated to further highlight the small targets while suppressing background noise. Experiments on the public datasets NUAA-SISRT, NUDT-SISRT and MDvsFA-cGan show that our proposed network outperforms the state-of-the-art methods in terms of detection accuracy and parameter efficiency.

高频金融数据背景下金融资产收益率序列普遍存在微观噪声结构,并且存在较为明显的重尾特征;同时,金融资产收益率的协方差矩阵具有高维性和稀疏性特征.基于预平均方法和Huber损失函数,采用收缩估计方法,得到高频金融数据背景下金融资产收益率的协方差矩阵的估计.模拟结果显示收缩估计方法有着较好的效果.此外,以中国A股市场资产的高频数据为样本进行实证分析,探究估计量在最小方差投资组合上的投资绩效.分析的结果显示:(1)预平均方法可以剔除绝大部分微观结构噪声对协方差矩阵估计的影响;Huber损失函数也可以减弱重尾现象对协方差矩阵估计的影响;(2)收缩估计量均能更好地估计总体协方差矩阵,并且在最小方差投资策略的比较中也拥有良好的投资绩效.