Generative artificial intelligence (GAI) is currently leading a fundamental transformation in higher education. This technological advancement offers a wide range of innovative applications, including the sophisticated generation of creative content, the implementation of adaptive learning methodologies, and highly personalized student support systems. Although the concept of trust in GAI (TIG) is a key determinant of students’ intention to use (IU), limited research has examined the influence of demographic factors on this relationship. This study investigates the effect of TIG on IU among Chinese university students, with gender and academic major as moderating variables. Using a cross-sectional survey design, data were collected from 481 undergraduate students across various academic majors at a public university in Henan Province, China, via an online questionnaire. The data were analyzed using PLS-SEM analysis to test the structural relationships and moderation effects. The results indicate a significant positive relationship between TIG and IU, with stronger effects observed among male students and those in Stem (Science, Technology, Engineering, and Mathematics) disciplines. The subsequent moderation analysis confirms that both gender and academic major significantly influence the TIG-IU relationship. These findings highlight TIG as a critical driver of GAI adoption in higher education and provide valuable insights for the development and implementation of targeted educational technology strategies. The study recommends that universities promote digital literacy, establish ethical GAI guidelines, and adopt discipline-sensitive pedagogical approaches to support the responsible and inclusive integration of GAI technologies in academic environments.
Emoji recommendation helps users enhance their communication on social media platforms by providing expressive and contextually relevant visual symbols. Emoji recommendation is language-dependent and remains an open challenge across languages. This study mainly investigates the impact of word-length and text-length on the performance of emoji recommendation for short Turkish texts. A dataset (B-D7) containing 158,201 tweets for the top 100 emojis grouped into 7 classes was created and used to generate 64 datasets based on combinations of word-lengths (2 to 5 characters) and tweet-lengths (1 to 16 words). The LinearSVC machine learning model was applied to each dataset. Additionally, two complementary experiments were conducted using the BiLSTM_wam deep learning model and the BERTurk transformer-based model on the datasets corresponding to the best and worst cases of LinearSVC. LinearSVC achieved its best accuracy (0.6502) and macro-averaged F1 score (0.5345) on the B-D7-wLt3-wt13 dataset, which includes tweets containing at least 13 words and only words with a minimum length of 3 characters, and its highest macro-averaged precision (0.5960) on B-D7-wLt2-wt8. The highest macro-averaged recall (0.5134) was shared between B-D7-wLt3-wt8 and B-D7-wLt3-wt13. The BERTurk model achieved its highest accuracy (0.7188), macro-averaged precision (0.6403), macro-averaged recall (0.5567), and macro-averaged F1 score (0.5691) on the BERTurk.B-D7-wLt3-wt8 dataset. LinearSVC and BERTurk consistently outperformed BiLSTM_wam across all corresponding dataset configurations and metrics. BERTurk outperformed LinearSVC across all dataset configuration-metric pairs, with only a single exception. Overall, the findings indicate that selecting appropriate word-length and text-length thresholds is crucial for maximizing emoji recommendation performance. Configurations that avoid aggressive filtering of shorter words while preserving sufficiently long texts provide richer contextual, semantic, and morphological information, leading to more reliable and generalizable results. This study may contribute to the development of more effective emoji recommendation systems for Turkish and other morphologically rich languages.
Within the coffee production chain, coffee husks are the most produced by-product, and a large proportion of the volume generated is not subjected to adequate treatment before disposal. This study aimed to evaluate the influence of geographical origin, altitude, and type of processing on the physical and chemical characterization of coffee husks (Coffea arabica) produced in Pernambuco, Brazil. The husks were coded based on information about the municipality, cultivar, handling, processing, and altitude. Analyses were carried out on the centesimal composition and physical-chemical characterization, and the data obtained were submitted to statistical analysis of variance (Anova), Tukey's test, and principal component analysis (PCA). The coffee husks showed significant differences in all the parameters evaluated. Samples from higher altitude regions (Triunfo) had higher values for bioactive compounds (caffeine and total phenolic compounds), while the fermentation process results in a relative concentration of insoluble carbohydrates. Detailed characterization can enable the transition of coffee husks from an underutilized by-product to a strategic technological raw material. The identification of distinct chemical profiles, influenced by altitude and fermentation process, provides the scientific basis necessary for its industrial valorization and the development of new products, and this reuse ensures sustainability in the coffee production chain and opens up new opportunities for producers.
This study introduces the HT Distribution (HTD), a novel one-parameter probability distribution derived as a special case of the two-parameter polynomial exponential distribution (Tped). The mathematical properties of HTD are explored, including its probability density function, survival function, hazard rate, moments, stochastic orderings, entropy, and stress-strength reliability. Actuarial applications are thoroughly examined, including mean excess loss, limited expected value, tail value at risk, and the effects of policy modifications such as deductibles, limits, and coinsurance. The HTD’s parameter is estimated using both the method of moments and maximum likelihood estimation, supported by simulation studies. The distribution’s effectiveness is validated through applications to two real-world datasets, where HTD outperforms several existing models based on standard information criteria.
Continuous exposure to dynamic environmental and mechanical stresses makes wind turbines (WTs) highly susceptible to component degradation and operational faults, thereby necessitating a reliable fault diagnosis framework to enable continuous monitoring and ensure safe, uninterrupted operation. Long Short-Term Memory (LSTM) networks are widely used for WT fault diagnosis due to their ability to capture temporal fault patterns; however, their diagnostic performance strongly depends on appropriate hyperparameter selection. To address this limitation, a Crested Porcupine Optimization (CPO) based adaptive hyperparameter tuning framework is proposed. The adaptive search capability of CPO enhances the LSTM’s ability to learn temporal fault signatures, thereby improving convergence, generalization, and diagnostic accuracy. To improve computational efficiency and mitigate overfitting, a ReliefF-PCA-based dimensionality reduction strategy was employed to identify the most informative Scada features. The proposed CPO-LSTM framework was validated using real world Scadadata collected from a 2 MW wind turbine. The classification accuracy and robustness of the proposed CPO-LSTM multi fault diagnostic framework were evaluated by comparing its performance against LSTM models optimized using conventional metaheuristic algorithms. Experimental results demonstrate that the proposed CPO-LSTM framework achieves a classification accuracy of 98.4% with strong robustness, indicating its suitability for practical WT condition monitoring applications.
