Against the backdrop of energy conservation, emission reduction and green low-carbon development, traditional syngas production processes are plagued by high energy consumption and excessive greenhouse gas emissions. In contrast, the photocatalytic CO2reduction technology, driven by solar energy, enables the conversion of CO2into syngas. It combines the values of environmental protection and resource recycling, thus becoming a current research hotspot. In the design of high-efficiency single-atom catalysts (SACs) for photocatalytic CO2reduction, the microenvironment design of single-atom metal sites is of crucial importance. Based on this, in this work, a series of Co-coordinated COF catalysts named Triazine-COF-Co-Cl were synthesized to regulate the Co-coordination microenvironment for enhancing syngas production via photocatalytic CO2reduction. Among them, the Triazine-COF-Co-AA catalyst with the best performance achieved a syngas production rate of 381.7 mmol g−1h−1, and the H2/CO molar ratio could be continuously adjusted in the range of 1-3, which is sufficient to cover the commonly used syngas ratio range in industry. This paper improves the performance of COF-based photocatalysts for syngas production via CO2reduction through the regulation of Co-coordination environment, and provides ideas and certain data support for the research on solving energy and environmental problems and reducing CO2emissions.

Show more

Emotion recognition (ER) poses a complex multi-class classification challenge, further complicated by significant class imbalances. In natural dialogue corpora, dominant emotions like neutral are prevalent, while minority emotions such as disgust and fear are notably scarce. This imbalance results in models consistently underperforming on less frequent categories. This paper investigates template-guided prompting as a method to improve long-tail emotion recognition using large language models (LLMs). We employ a unified evaluation framework on the MELD dataset to compare various methods: supervised baselines (TextCNN, BiLSTM), a fine-tuned pre-trained model (BERT-base), and training-free LLM inference (DeepSeek) using three structured prompt templates in both zero-shot and few-shot scenarios (K=1, 3, 5, 10). Our findings demonstrate that template-guided LLM prompting achieves the highest overall performance (Acc=0.6573, Macro-F1=0.5268) and significantly enhances minority-class F1 scores compared to all supervised baselines, without requiring parameter updates. A detailed analysis of hard-sample errors shows that 16.9% of test instances are misclassified by all five models, with minority emotions having hard-sample rates up to 48%. This bias remains even with balanced downsampling (Pearson r=0.986) and is linked to a systematic prediction bias toward the neutral class. These results imply that the difficulties in long-tail ER arise from intrinsic semantic ambiguity rather than just data imbalance, and that structured prompting offers a practical and effective solution for achieving more balanced emotion recognition.

Show more

This study aims to improve the early prediction of academic paper impact. To this end, a dual-pathway dynamic heterogeneous graph neural network model is proposed. The model constructs a temporal heterogeneous graph consisting of multiple types of nodes, including papers, authors, institutions, and journals. Citation and non-citation pathways are designed to model knowledge diffusion relations and social-semantic associations, respectively. In addition, a life-cycle-aware mechanism is introduced to capture feature variations across the emerging, growing, and mature stages of papers. On this basis, the prediction outcome is further disentangled into three independent components, namely diffusion effect, social bias effect, and intrinsic value effect, thereby enhancing the interpretability of the model. Experimental results on datasets from multiple disciplines, including computer science, chemistry, and psychology, demonstrate that the proposed method outperforms existing mainstream models in terms of mean absolute log error and log-transformed coefficient of determination. It is especially more accurate and stable in cold-start scenarios. The results indicate that the proposed method can provide effective support for the early identification of high-potential papers.

Show more

In the wave of Industry 4.0, industrial Internet of Things (iot) devices are increasingly widely used in manufacturing, energy, chemical and other fields. The scale and operational complexity of these devices are constantly rising, and their stable operation is crucial for production efficiency and safety. To make up for the shortcomings of existing algorithms in long time series feature extraction and complex association mining, this paper proposes the Transformer-BiGRU classification and regression algorithm, and first conducts correlation analysis and violin plot analysis. Experiments show that the core evaluation indicators of the classification algorithm have significant advantages. The accuracy rate, recall rate, and precision rate all reach 84%, 84%, and 86% respectively. The F1 value is 83%, all higher than all comparison machine learning algorithms. The AUC reaches 93%, although slightly lower than CatBoost's 94%, it is higher than other algorithms such as Random Forest. Strong generalization and category discrimination capabilities; The MSE of the regression algorithm is 16.598, the RMSE is 4.074, and the MAE is 2.615, all of which are the lowest. The MAPE is 51.468, which is in a relatively low range. The R² reaches 0.442, which is significantly better than the traditional algorithm. This algorithm integrates the global dependency capture capability of Transformer with the temporal feature extraction advantages of BiGRU, providing a reliable solution for the precise analysis and prediction of the operating status of industrial Internet of Things devices, which is of great significance for ensuring the efficiency and safety of industrial production.

Show more