Sparse representation face recognition does not have the ability to distinguish high-dimensional data and can not describe the distance information between data. The KTPTSR method proposed in this article uses Gaussian kernel function to map data to high-dimensional space, and uses two-stage sparse representation method to classify training data. Tests on ORL and YALE facial databases show that KTPTSR has a better average recognition rate than the classic algorithms, and has stronger robustness to changes in the number of neighbors. The kernel operation also reduces computational overhead. This method effectively combines the advantages of kernel methods and sparse representation, providing a reliable improvement solution for face recognition in complex scenes.
Guidelines for professional degree theses require automatically generated feedback that is both linguistically accurate and consistent with program-specific structural, methodological, and regulatory requirements. This study developed a retrieval-augmented educational artificial intelligence system that assigns confidence scores to four thesis types, retrieves type-specific normative requirements, and applies a confidence boundary mechanism to select single-type or dual-type feedback. A controlled pilot experiment compared four system variants on 50 expert-annotated Education Master’s theses: a baseline large language model, a prompt-enhanced variant, a retrieval-augmented variant, and the full system that adds the confidence boundary mechanism to retrieval. Under a strict unique-primary-label rule, the baseline and the prompt-enhanced variant reached 64.0 % and 46.0 % accuracy, respectively. The independently executed retrieval-augmented variant and the full system reached 80.0 % and 82.0 %, and this paired difference was not statistically significant. A routing-only paired reanalysis that reused the same confidence scores returned 82.0 % accuracy with the boundary mechanism both disabled and enabled, confirming that the mechanism governs feedback routing rather than the choice of the top-ranked thesis type. With the mechanism enabled, boundary detection reached 90.0 % precision, 87.1 % recall, and an F1-score of 88.5 %. Complete dual-type coverage was 61.3 %, and partial candidate coverage reached 87.1 %. Stratified-bootstrap confidence intervals and macro-averaged scores were reported to reflect the small, imbalanced sample. The results indicate that retrieval grounding supports the assignment of a primary thesis type, whereas boundary-aware routing reduces overconfident single-type guidance for ambiguous thesis topics.
Edge cloud collaborative systems face critical challenges in deploying virtual network functions with dynamic adaptability and cost efficiency due to the heterogeneity of Virtual Machine-Virtual Network Function (VM-VNF) and Container Technology-Virtual Network Function (CT-VNF) resources and high environmental dynamics. This study proposes a Long Short-Term Memory Network (LSTM)-enhanced Actor-Critic framework for VNF deployment optimization. Experimental results demonstrate that the framework achieves a deployment success rate of 97.9 %, mapping success rate of 94.8 %, average latency of 362 ms, and optimal solution fitting degree of 0.93, significantly outperforming baseline methods. With average costs 30–40 % lower than comparison frameworks under varying request loads, this study provides an effective solution for service providers to implement cost-minimization management strategies in edge cloud environments. The innovation of this study lies in designing a temporally coherent dual-network architecture rather than a simple combination of existing modules. In the actor module, LSTM encodes sequential request traces to enable anticipatory deployment decisions. In the critic module, LSTM-empowered approximation is coupled with Lagrangian relaxation to evaluate long-term constraint satisfaction under mixed VM-CT deployment. This design mitigates the high variance and slow convergence inherent in conventional Actor-Critic methods under dynamic network conditions.
Highway construction projects, characterized by large investments, extended durations, and scattered work sites, require accurate and dynamic progress monitoring to ensure cost control and safe execution. Traditional manual inspection methods suffer from inefficiency, data latency, and subjectivity, making them inadequate for modern intelligent transportation management. To address these limitations, this study proposes a novel monitoring method integrating unmanned aerial vehicle (UAV) remote sensing imagery with feature-space matching. The proposed framework involves collecting sequential site images via UAV oblique photography, generating high-precision 3D reality models through aerial triangulation and 3D reconstruction, determining optimal image segmentation scales using a multi-scale segmentation algorithm combined with the ESP2 tool, extracting construction targets via color-feature threshold segmentation and registration with BIM-based virtual models, and quantifying progress deviations through Boolean operations between 3D reality models and BIM envelopes. The primary objective is to establish a closed-loop workflow from data acquisition to progress quantification, enhancing monitoring accuracy and operational efficiency. Experimental results on a 30 km highway section over 42 days show a construction progress deviation of only 2.2 %, significantly lower than conventional methods. The proposed method provides an objective, quantitative, and visual monitoring solution, demonstrating high accuracy, stability, and practical potential for improving schedule management in highway construction projects.
This study investigates the influence of input window size and ensemble size on the forecasting performance of Decision Tree-based ensemble methods in one-step-ahead financial return prediction. Nineteen financial instruments spanning stocks, indices, commodities, and cryptocurrencies are evaluated under a chronological train-validation-test framework. Daily closing prices are transformed into log returns, and models are trained across rolling windows of 5, 10, 20, and 60 observations. The methods compared include Naive forecasting, Linear Regression, a standalone Decision Tree, Bagging-DT, Boosting-DT, and Stacking-DT, with ensemble configurations selected based on validation MAE. Linear Regression achieves the lowest average magnitude error and the best overall rank, while Bagging-DT and Boosting-DT substantially outperform both Naive forecasting and the standalone Decision Tree. The validation-selected ensemble reduces MAE across all 19 instruments, yielding an average out-of-sample improvement of 31.89 % relative to the single tree. A test-best diagnostic comparison suggests a higher improvement ceiling of 35.10 %, though this figure serves solely as an upper-bound reference. A 20-day window delivers the strongest average error performance, while a 60-day window and 100-estimator ensembles are most frequently selected at the individual instrument level. Larger ensembles incur additional computational cost without producing consistent accuracy gains. Friedman and Wilcoxon tests confirm significant performance differences among models, and Diebold-Mariano tests indicate that ensemble methods consistently outperform both Naive forecasting and the standalone Decision Tree. Overall, Decision Tree ensembles offer robust improvements over persistence-based and single-tree forecasting, though their advantage relative to Linear Regression remains instrument-dependent.
