Exploring a quantitative assessment system for the popularity of digital media art has become a central issue linking art creation, technology application and market communication. Aiming at the bottleneck of the current algorithm, this paper proposes a LSTM algorithm based on multi-head attention mechanism optimization. The study began with a correlation analysis. The results showed that the number of interactive elements had the strongest correlation with popularity, and the absolute value of correlation coefficient reached 0.556887. Variables such as creation time, number of colors, and complexity score also have some correlation with popularity. It can be seen that the time invested in creation, the richness of color and the complexity of works will affect the popularity of works to a certain extent. Taking decision tree, random forest, CatBoost, AdaBoost, and XGBoost as comparative experimental objects, our model performed best in terms of accuracy, recall, accuracy, F1, and AUC in various indicators. Its accuracy was 0.855, which was higher than decision tree (0.709), random forest (0.803), CatBoost (0.786), AdaBoost (0.778) and XGBoost (0.744), and the overall classification accuracy was the highest. Recall and accuracy of 0.855 and 0.856, respectively, are also ahead of other models. It performs better at identifying positive samples and predicting the proportion of actual positive samples among positive samples. The F1 value of 0.855 is also higher than that of other models, and the ability to take into account both accuracy and recall is stronger. The AUC reached 0.904, surpassing random forest's 0.875 and CatBoost's 0.876, demonstrating the best ability to differentiate between positive and negative samples. In contrast, the other models are slightly inferior in various metrics, especially the overall performance of the decision tree and XGBoost is significantly lower than ours. This research outcome provides a more effective method for quantitatively assessing the popularity of digital media art. It not only provides a direction for the integration of artistic creation and technology application, but also provides a scientific basis for the formulation of market communication strategy.

Retrieval Enhanced Generation (RAG) system improves the accuracy and reliability of content Generation by retrieving external knowledge, and has been widely used in intelligent question answering, knowledge assistant and other fields. However, its core performance depends on the quality of the retrieval stage, and the relevance and factual consistency of the retrieval results directly determine the validity of the generated content. However, factors such as query complexity, document noise, and domain differences in real-world scenarios can easily lead to fluctuations in retrieval quality. Traditional manual evaluation is costly and outdated, and it is difficult to meet the real-time optimization requirements. At the same time, the existing models have limitations in complex feature fusion and parameter optimization. Therefore, this paper proposes a retrieval quality prediction model combining lizard optimization algorithm (HLOA), convolutional neural network (CNN), and bidirectional gated recursive unit (BIGRU). The correlation analysis shows that there is a strong positive correlation between the search level and the search usefulness score, that is, the higher the search level, the better the search usefulness score; There is a strong negative correlation between query complexity and retrieval usefulness score, which means that the higher the query complexity, the lower the retrieval usefulness score. Comparing this model with 9 models including decision tree, random forest Adaboost, gradient boosting tree, ExtraTrees, CatBoost, XGBoost, LightGBM, and KNN, it is shown that their performance is better overall: MSE (28.617), RMSE (5.349), MAE (4.401), and MAPE (17.355) are the lowest, while R² (0.952) is the highest. This study provides an effective solution for accurate prediction and real-time optimization of RAG system retrieval quality, which helps to enhance the application value of RAG technology in practical scenarios.

With the advancement of generative artificial intelligence, AI-based text generation has been increasingly applied to the domain of screenplay writing, raising critical questions about whether algorithmic storytelling can embody literariness, cultural expression, and philosophical depth. This study uses Life of Pi as a case and constructs two AI-generated screenplay samples (theme-driven and adaptation-driven) to compare systematically with Ang Lee’s directorial version. Methods include narrative structure modeling, thematic weight analysis, symbolic language density computation, and philosophical abstraction measurement. A multidimensional comparison across narrative coherence, thematic focus, linguistic tension, and cultural depth is conducted, complemented by blind expert interviews involving five specialists to evaluate literary expressiveness from a humanistic perspective. The results show that while AI scripts perform well in structural control and thematic identification, they lag behind auteur-driven screenplays in philosophical abstraction, symbolic system construction, and aesthetic articulation. The study concludes that current AI systems are not yet capable of independently producing screenplays with humanistic depth but can function as effective tools in generating genre-oriented drafts.

In view of the profound impact of digital transformation on public services in China, there is an emerging challenge of balancing efficiency and equity in public service reform. On one hand, the efficiency of public services has been greatly improved by means of open government data and the development of smart platforms. On the other hand, efficiency-oriented mechanisms may undermine the security of the institution and cause risks to social trust. How to realize the efficient optimization while guaranteeing the security of the institution has become a crucial issue for current public governance. This study has put forward a methodology that integrates the constraint of institutional security with interpretable reinforcement learning. Through incorporating publicly accessible government service data into the construction of the state space and reward function, considering the regulations of the institution within the model and an interpretable model describing the decision process, this methodology can make the optimal tradeoff among efficiency optimization, institutional regulations, and interpretability. The empirical results demonstrate the superiority of this methodology over traditional ones in terms of saving processing time, optimizing user satisfaction and ensuring the provision of services for vulnerable groups. In addition, this study not only offers a technical roadmap for public service process optimization but also offers assistance for methodological innovation in institutional innovation and e-governance.

In this paper, we present a framework that uses LLMs to predict scores of classroom feedback that employs dual alignment mechanisms to guarantee interpretability and fairness. Specifically, we tackle the decades-old problem of the black-box nature of automatic scoring by leveraging semantic alignment through attention regularization and pedagogical alignment through rubric fine-tuning. Data were collected from more than 65,000 classroom feedback responses in secondary and higher education settings from three countries. This resulted in the collection of more than 7.3 million words of analyzed text. Preprocessing procedures were done to ensure ethical adherence and preserve discourse structure. Experimentation on the dataset yielded significant increases in terms of prediction accuracy, robustness to rubric changes, and interpretation results. Specifically, our proposed framework outperforms baselines by 21.3% and 27.4% in terms of root mean square error and rubric coherence, respectively. Statistical testing confirmed improvements for all rubric dimensions where the effect size is either medium or large (Cohen's d = 0.63–0.87). Our teacher survey also found that 82% of teachers trusted model output more, whereas confirmatory factor analysis found three-factor structure of trust, pedagogical meaning, and usability with high internal reliability (Cronbach's α = 0.92). Our research shows that interpretability does not necessarily come at the cost of accuracy.

The GBA region has become an attractive area for the development of cross-border medical tourism, with the help of online platforms uniting patients, hospitals and facilitators across different territories. However, fast-growing trends in data regulation and differing governance systems bring uncertainty as to how such ecosystems might develop. To that end, this paper proposes a regulation-aware agent-based simulation (ABM) modeling interaction of hospitals, facilitators, patients and regulatory bodies within a period of five years. The presented ABM framework introduces the elements of heterogeneity of agents' preferences, bounded rationality and adaptive learning in changing regulations. Based on Monte Carlo simulations across scenarios of lenient, phased and strict regulation, we establish non-linear thresholds in ecosystem emergence. Our results show that phased regulation creates a critical mass in the ecosystem by month 26, resulting in 2.1 times higher amount of cross-border patients than under lenient regulation and not leading to ecosystem collapse like in case of strict regulation. Sensitivity analysis performed with Sobol indices allows us to establish the most important input variables: the cost of data localization and the cost of cross-border consent with the importance of 61.2% and 22.5% of the output variance, respectively. In addition, two functional models are introduced: the model of update equation of facilitators' strategies based on reinforcement learning and variance decomposition model for estimation of parameters' effects. Empirical calibration with telemedicine adoption rates and validation based on out-of-sample 2024-Q4 statistics confirm the validity of the proposed model.

In order to meet the needs of enterprises for accurate classification of massive text data in the process of digital transformation, this paper applies the optimized text classification model combining prompt engineering and AIGC to enterprise text classification task. And choose classical decision trees, mainstream ensemble learning models (Random Forest, AdaBoost, GBDT, ExtraTrees) and high-performance gradient enhancement model XGBoost as comparison models. The performance of the model was evaluated by five metrics: mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²). The experimental results show that the MSE of our model is 14.153, which is significantly lower than all comparison models and about 13.4% lower than the suboptimal AdaBoost (16.352). Its RMSE (3.762), MAE (3.069), and MAPE (5.029) are also the smallest of all models, decreasing by 7.0%, 3.7%, and 2.2% compared to the corresponding metrics of AdaBoost (4.044, 3.186, 5.142), respectively, indicating that the model has smaller prediction bias and better accuracy in category estimation. Meanwhile, the R² value of our model reaches 0.826, which is higher than the comparative models such as Random Forest (0.74), GBDT (0.783), and XGBoost (0.732). It can explain 82.6% of text category changes, capturing the mapping relationship between text features and categories more accurately. The above results validate the effectiveness of the collaborative optimization strategy of cue engineering and AIGC-cue engineering can guide the model to focus on key semantic features of text to reduce feature extraction bias, while AIGC can supplement high-quality text samples or enhance feature expression dimensions to alleviate data sparsity. The combination of the two significantly improves the prediction accuracy and stability of the text classification model.

Reliable online monitoring of Gas Tungsten Arc Welding (GTAW) is difficult because visual, electrical, and acoustic observations each describe only part of the welding process. In this work, we propose a compact multimodal fusion network with frequency-dynamic convolution (FDConv) to improve weld-state recognition under real-time constraints. The network applies modality-specific spectral enhancement before intermediate fusion, enabling effective integration of synchronized arc current/voltage signals, acoustic emission (AE) spectrograms, and infrared (IR) weld-pool images. On a balanced GTAW dataset, the proposed method achieves an F1-score improvement of about 4–5 percentage points over a tuned CNN–LSTM fusion baseline, while maintaining an added latency of no more than 100 ms at a 10 Hz decision rate. The experimental results show that emphasizing informative frequency components prior to fusion helps retain defect-sensitive patterns and yields more stable recognition performance. These findings support the use of frequency-aware multimodal learning for real-time GTAW quality monitoring.

Unplanned stoppages in semiconductor tools remain a persistent limiter of throughput and yield, a situation partly sustained by monitors that rely on dense labels or rule sets that do not travel well across recipes and tools. We study a digital-twin-driven framework that learns a compact health representation from multiscale telemetry by self-supervised objectives and regularizes it with differentiable constraints drawn from mass balance, thermal–RF coupling, and vacuum dynamics; anomaly evidence is then fused with process, environmental, and maintenance logs so that alerts arrive with context and with a plausible operational hypothesis. Orchestrated with DolphinScheduler or Airflow, the pipeline coordinates ingestion, training, streaming inference, lineage, and review to align analytics with change control and auditability. Development was deliberately iterative rather than linear: label sparsity and timestamp drift pushed us toward cycle-aware alignment; twin mis-specification in edge regimes required residual diagnostics and parameter re-estimation; population shift prompted conformal calibration and sequential testing. On production-like etch and deposition traces, we observe earlier detection under fixed alert budgets and extensions in lead time that appear to improve MTBF and OEE to some extent, together with indications of lower service effort and energy use. Alternative explanations, including facility subsystems or undocumented operator interventions, cannot be excluded, which suggests that further research is needed on causal attribution, cross-site transfer, and adaptive twin updating.

Degenerative retinal diseases are heterogeneous, progressive diseases with a molecular complexity that includes many different aspects of oxidative stress regulation which have not yet been fully clarified. In order to solve the problem of limited interpretability of the existing imaging-based model approaches, this paper offers a framework of an interpretable graph-biochemical pathway model which combines transcriptomic, retinal OCT image data and biological pathways information to model the dynamic behavior of the NRF2-ROS regulatory loop. The proposed framework reaches 0.923 accuracy and 0.945 AUC in five-fold cross validation experiment on the GSE29801 dataset and surpasses several machine learning and GNN-based baseline algorithms. Twenty-three key regulatory nodes and fifteen significant edges have been identified in this work, and the scores of pathway activation are well correlated with the visual function measures.