The encode network traffic makes it difficult to perform successful and dynamic classification. This paper will present the use of a hybrid model to be used with the publicly available QUIC dataset to classify VPN and non-VPN encrypted traffic based on a Deep Convolutional Neural Network (DCNN) and Long Short-Term Memory (LSTM) network, which is optimized by the Artificial Bee Colony (ABC) and Genetic Algorithm (GA). The method involves multi-angle processing - preprocessing, Min-Max normalization, and features selection of with correlation analysis, Fisher Score, and mutual information to obtain a tiny, but meaningful feature set (Size, Batch Cache, Delta Previous Packet). The chosen features are translated to 2D tensors through a sliding time window of consecutive packets, which allows the spatio-temporal DCNN+LSTM architecture to represent the level of intra- and inter-packet feature associations as well as inter- and intra-packet time dynamics. The disadvantages of single-optimization are overcome using a dual metaheuristic optimization strategy again whereby the work of the global hyperparameter exploration is done using ABC and the structural optimization is done using GA. The imbalance of classes is reduced with weighted loss functions and stratified data division. The accuracy of the model is 99.66% with 0.994 ROC-AUC and 0.987 PR-AUC and its MCC is 0.963 which is even greater than that of the traditional classifiers (Decision Tree, Random Forest, SVM, KNN), individual deep-learning models (CNN, LSTM), and image-based FlowPic method. Three-quarters of stratified cross-validation marks the case of consistent generalization (99.53% ± 0.09% mean accuracy), and an ablation study confirms the value of any one of the components. The findings prove that the presented framework can be applied to monitor the network traffic on encrypted networks which are security-sensitive and in real-time.

This study examines the role of generative artificial intelligence (GenAI) in cybersecurity threat detection, focusing on its usefulness in workflows that support human decision-making. Experiments were performed on the BODMAS dataset (134,435 samples) and a smaller exploratory subset of UNSW-NB15. State-of-the-art machine learning (ML) classifiers were compared with a zero-shot large language model (LLM) using standard classification metrics, while also considering latency, cost, and hallucination risk. ML classifiers consistently outperformed the LLM-based system on standard detection metrics. However, the LLM showed value in cases of ambiguity, where it could provide short plain-language explanations, organize alert-related context, and generate initial interpretations for instances that did not match learned classes. GenAI is unlikely to replace ML-based detection methods, but it can provide useful interpretive support for ambiguous or unfamiliar alerts. A hybrid pipeline is therefore proposed, in which ML handles high-confidence and time-sensitive decisions, while the LLM is used selectively for low-confidence cases or when explanatory support is needed. Human oversight remains necessary to address hallucination risk and ensure reliability.

Vehicle insurance claim severity modeling requires accurate and interpretable methods that can handle skewed and heterogeneous loss data. This study provides a structured empirical comparison between classical parametric regression models and tree-based ensemble learning approaches for predicting claim size conditional on claim occurrence. The analysis is conducted within a cross-sectional conditional severity framework using real-world motor insurance data. We implement and compare ordinary least squares (OLS), a Tweedie generalized linear model (GLM), and three ensemble methods: bagging, random forests (RFs), and gradient boosting. Model performance is evaluated using out-of-sample root mean square error (RMSE), and variable importance measures assess the relative contribution of predictors. The results indicate that tree-based ensemble methods achieve modest improvements in predictive accuracy relative to classical parametric models. The Tweedie GLM remains a competitive, flexible parametric benchmark for skewed positive claim amounts. Variable importance analysis consistently identifies premium and insured value as key determinants of claim severity. Overall, the findings suggest that ensemble learning methods can complement traditional actuarial models, offering additional flexibility in capturing non-linear effects while maintaining comparable predictive performance in moderate-complexity severity data.

IntroductionDigital platforms have increasingly reshaped the ways in which traditional craft cultures are produced, circulated, and interpreted. While prior research has examined digital heritage broadly, limited attention has been paid to how platform-based dissemination transforms ceramic culture in historically significant craft centers such as Jingdezhen.MethodsThis study adopts a qualitative research design, combining semi-structured interviews with 32 ceramic practitioners and digital ethnography of 58 ceramic-related livestreaming sessions on Douyin.ResultsThe findings reveal three key dynamics: (1) the reconfiguration of craft authority through platform visibility; (2) the emergence of hybrid artisan–educator–entrepreneur identities; and (3) persistent tensions between cultural authenticity and commercial logic in platform-mediated environments.DiscussionBy integrating cultural ecology and platform ecosystem theory, this study contributes to scholarship on digital heritage and provides practical insights for cultural practitioners and heritage institutions navigating digital platform ecosystems.

Improving software efficiency is crucial for maintenance, but pinpointing runtime bottlenecks becomes increasingly difficult as systems expand. Traditional dynamic profiling tools require full build-execution cycles, creating significant latency that impedes agile development. To address this, we introduce GFTrans, a static analysis framework that predicts c program performance without execution. GFTrans utilizes a Transformer architecture with a novel “anchor-based embedding” technique to integrate control flow and data dependencies into a unified sequence. Additionally, a dynamic gating mechanism fuses these semantic representations with 16 handcrafted statistical features to comprehensively capture code complexity. Evaluated on a dataset of real-world GitHub c functions with high-precision runtime labels, GFTrans outperforms baseline models like Random Forest and Code2Vec, achieving 78.64% accuracy. The system identifies potential bottlenecks in milliseconds, enabling developers to perform optimization effectively during the coding phase.

Anomaly detection in video surveillance remains a challenging problem due to complex human behaviors, temporal variability, and limited annotated data. This study proposes an optimized spatiotemporal deep learning (DL) framework that integrates a Convolutional Neural Network (CNN) for spatial feature extraction with a Long Short-Term Memory (LSTM) network for temporal dependency modeling. The CNN processes frame-level appearance information, while the LSTM captures sequential motion patterns across video frames, enabling effective representation of anomalous activities. Hyperparameter optimization and regularization strategies are employed to improve convergence stability and generalization performance. The proposed model is evaluated on the DCSASS surveillance dataset and the experimental results demonstrate that the optimized CNN-LSTM framework achieves an accuracy of 98.1%, with consistently high precision, recall, and F1-score across 3-fold, 5-fold, and 10-fold cross-validation settings. Comparative analysis shows that the proposed method outperforms conventional machine learning models and recent deep learning baselines, highlighting its effectiveness and robustness for practical video-based anomaly detection in surveillance environments.

High-dimensional streaming data implementations commonly utilize online streaming feature selection (OSFS) techniques. In practice, however, incomplete data due to equipment failures and technical constraints often poses a significant challenge. Online Sparse Streaming Feature Selection (OS2FS) tackles this issue by performing missing data imputation via latent factor analysis. Nevertheless, existing OS2FS approaches exhibit considerable limitations in feature evaluation, resulting in degraded performance. To address these shortcomings, this paper introduces a novel genetic algorithm-based online sparse streaming feature selection (GA-OS2FS) in data streams, which integrates two key innovations: (1) imputation of missing values using a latent factor analysis model, and (2) application of genetic algorithm to assess feature importance. Comprehensive experiments conducted on six real-world datasets show that GA-OS2FS surpasses state-of-the-art OSFS and OS2FS methods, consistently attaining higher accuracy through the selection of optimal feature subsets.

Spatio-temporal crime prediction is crucial for optimizing police resource allocation but faces challenges including data sparsity, which hinders models from extracting effective patterns and limits robustness-and the underutilization of cross-type crime co-occurrence correlations. To address these issues, we propose a transfer learning approach that explores underlying cross-type relationships, enabling the sharing of spatio-temporal features across crime types and alleviating data sparsity. An adaptive weight updating mechanism is incorporated to enhance the perception of distinct crime categories, while the impacts of points of interest (POIs), meteorological factors, and other features are also analyzed. Experiments on real-world data from a Chinese city show that our model comprehensively captures latent features across crime types, thereby enhancing predictive performance and robustness, particularly for crime types with sparse data. Moreover, it effectively incorporates environmental features, further improving crime prediction performance.

Long-term fairness in sequential decision-making is critical yet challenging, as decisions at each time step influence future opportunities and outcomes, potentially exacerbating existing disparities over time. While existing methods primarily achieve fairness by directly adjusting decision models, in this work, we study a complementary perspective based on sequential algorithmic recourse, in which fairness is pursued through actionable interventions for individuals. We introduce Sequential Causal Algorithmic Recourse for Fairness (SCARF), a causally grounded framework that generates temporally coherent recourse trajectories by integrating structural causal modeling with sequential generative modeling. By explicitly incorporating both short-term and long-term fairness constraints, as well as practical budget limitations, SCARF generates personalized recourse plans that effectively mitigate disparities over multiple decision cycles. Through experiments on synthetic and semi-synthetic datasets, we empirically examine how different recourse strategies influence fairness dynamics over time, illustrating the trade-offs between short-term and long-term fairness under sequential interventions. The results demonstrate that SCARF provides a practical and informative framework for analyzing long-term fairness in dynamic decision-making settings.

IntroductionThe rise of contactless payment has made quick response (QR) code-mobile payment (QR-MP) platform increasingly popular among mobile financial service (MFS) users, especially in emerging economies. It has been demonstrated that the ongoing use of QR payments can significantly drive the growth of emerging economies. However, despite its importance, the continued use of this technology has not been satisfactory. Thus, this study seeks to explore the modified Unified Theory of Acceptance and Use of Technology 2 (UTAUT2) model, including four additional constructs: amotivation (AM), alternative attractiveness (AA), QR transaction anxiety (QTA), and transaction convenience (TC) to examine the MFS users' sustained usage of QR payment.MethodsData were collected from 247 MFS users in Bangladesh using an online survey and analyzed through SEM-PLS and non-linear analysis of IPMA.ResultsThe research findings reveal that effort expectancy is the most influential factor, and that both moderator factors, QTA and TC, are significant. However, social influence and hedonic motivation were found to be insignificant. Furthermore, our extended research model explains 76.5% of the variance in CINT without the moderation effect.DiscussionThe IPMA findings help to find the best-performing variables and provide practical insights for this study. Theoretical and managerial implications are provided to enrich the existing literature on the study of information technology, indicating how MFS providers in developing countries can retain their existing users.