The effectiveness of static trading strategies is often undermined when markets undergo non-stationary concept drift, which, in turn, precipitates increased return uncertainty and persistent performance deterioration. To address this challenge, we present a drift-aware dynamic trading strategy portfolio framework for streaming financial environments, incorporating ADWIN-based drift detection and a semi-supervised classification mechanism. The proposed framework operates through two stages. During the offline stage, initial trading strategy portfolios are constructed by applying memetic algorithm to historical data, thereby generating a baseline trading strategy portfolio. In the online stage, ADWIN continually monitors streaming market indicators to identify abrupt changes in statistical behavior. Once drift is detected, a semi-supervised classifier is updated using a combination of limited labeled data and pseudo-labeled streaming instances, enabling it to reassess current market regimes and guide adaptive adjustments to strategy parameters or partial portfolio replacements. Experiments conducted using Taiwan stock market data demonstrate that the proposed framework delivers substantial performance improvements over baseline strategies.
This study develops a pilot-type multimodal framework for predicting purchase intention in digital environments by integrating facial-expression analysis and eye-movement dynamics. While prior research has often examined facial expressions or visual attention separately, purchase intention is likely to emerge from the interaction between affective response and visual-cognitive processing. To explore this possibility, the present study adopted a small-scale within-subject pilot design involving three Taiwanese participants who browsed four digital platform conditions: traditional online shopping websites, Facebook, Instagram, and TikTok. Facial Action Units (AUs) and eye-related variables were analyzed in relation to post hoc judgments of purchase intention. The findings suggest that purchase intention cannot be sufficiently explained by facial expressions alone. Instead, it is better understood as a multimodal behavioral state characterized by platform-specific combinations of facial activity and eye-movement patterns, including scanning behavior, attentional engagement, and depth-related visual focus. The study contributes by constructing an exploratory pilot-type analytical framework that integrates facial and ocular indicators for the behavioral assessment of purchase intention across digital platforms. This framework may serve as a foundation for future large-scale research in digital marketing and consumer behavior.
Airports are complex built environments where passengers rely on effective wayfinding systems to navigate time-sensitive processes. However, traditional navigation methods such as signage and digital displays often become insufficient in visually dense and spatially complex terminals. Kuala Lumpur International Airport 2 (KLIA2) exemplifies these challenges due to its large-scale layout and high passenger volume. This study presents AR-Port, an augmented reality (AR)-based indoor navigation system designed to support intuitive wayfinding through spatially anchored visual guidance. The system integrates on-device environment scanning, waypoint-based path definition, and AR rendering to provide real-time navigation cues aligned with the user’s physical surroundings. A user acceptance evaluation involving 86 participants was conducted to assess perceived navigation difficulty and openness toward AR-based guidance. Results indicate that a majority of participants experience challenges with existing navigation tools and show positive receptiveness toward mobile and AR-assisted navigation. While the evaluation focuses on user perception rather than system performance, the findings highlight the potential of AR-based navigation to enhance passenger experience and reduce cognitive load in complex airport environments. Future work will include empirical benchmarking of localization accuracy, anchor stability, and navigation efficiency in real-world settings.
With the proliferation of obesogenic environments for companion dogs, the objective measurement of the Body Condition Score (BCS) has become increasingly critical as a key health indicator. To overcome the limitations of conventional Convolutional Neural Network (CNN)-based unimodal approaches, this study proposes Meta-Cobi, a Vision-Language Model (VLM)-based framework that integrates multi-view visual data with biological metadata. The framework incorporates the Convolutional Block Attention Module (CBAM) to maximize the density of captured anatomical features and utilizes a Learned Query Pooling mechanism to globally compress information gathered from 11 distinct viewpoints. Furthermore, biological metadata such as breed and age is integrated into structured prompts, while a task-specific special token strategy is employed to enhance inference precision. Experimental results demonstrate that Meta-Cobi achieves superior Macro-F1 score and robustness compared to baseline models, notably maintaining high discriminatory performance even in data-sparse classes. This research expands the scope of AI-driven BCS assessment into the veterinary domain and contributes to the establishment of data-driven precision healthcare systems for companion animals.
Higher Education Institutions in Latin America face several challenges, the most significant being digital transformation and the efficient management of intellectual capital. Among the obstacles they encounter are the persistence of “data silos” that fragment institutional information and hinder strategic decision-making. Therefore, this article presents a conceptual work-in-progress proposal for an interoperable data governance framework. While the overall methodology encompasses three phases, the current focus is on defining the fundamental architecture designed to enable and strengthen Knowledge Management in public university environments.
Fraud and anomalous activities in crowdfunding campaigns have become increasingly critical, as misleading content can undermine user trust and platform sustainability. This research proposes a multimodal framework to detect image-text inconsistency by integrating CLIP-based similarity, OCR-extracted textual features, and project-level feature aggregation. By jointly analyzing visual and textual information, the proposed approach captures semantic misalignment that is often overlooked in prior research focusing on textual data in fraudulent crowdfunding detection. Three classifiers: Logistic Regression, Random Forest, and XGBoost were evaluated with imbalance handling techniques, including SMOTE, class weighting and hybrid strategies. Experimental results showed that XGBoost with SMOTE achieved the best performance. Sensitivity analysis further indicated that a 20
This paper develops a conceptual and empirically illustrated framework for constructing a service ecosystem through the metaverse by integrating servitization and decentralized autonomous organizations (DAOs). We synthesize top-journal research on service ecosystems, digital servitization, metaverse platform ecosystems, and DAO governance to identify a central challenge: the metaverse expands the design space for service experiences and resource integration, yet its ecosystem-level coordination demands governance mechanisms that can scale across heterogeneous actors, assets, and jurisdictions. We propose a layered model in which servitization modularizes and operationalizes metaverse services, while DAOs instantiate programmable, auditable ecosystem governance (incentives, rules, and collective decision rights). A Japanese manufacturing case illustrates feasibility via digital twin, enabled service delivery and platform partnerships, and clarifies implementation requirements for DAO-based ecosystem coordination.
This study proposes a dynamic decision support framework for feed design in pig production by integrating biological states and future outcomes. Lactation feed design has been suggested to be associated not only with current performance but also with subsequent performance (e.g., weaning-to-service interval, WSI). However, practical decision-making is typically based on short-term indicators, and dynamic evaluation remains underdeveloped. In this study, litter average daily gain (ADG) is treated as a state variable linking feed design to future outcomes. Random forest models are used to estimate state formation and future performance, and a two-period simulation is conducted to evaluate dynamic economic outcomes under discounting. The results indicate nonlinear relationships between feed design, ADG, and subsequent performance. Although exact optimal points could not be determined reliably, broad state-dependent tendencies remained observable across parity groups. Differences were also observed between short-term and dynamically evaluated recommended designs. These findings suggest that feed design should be interpreted as a state-dependent and exploratory decision problem rather than a deterministic optimization problem. The proposed framework demonstrates how observational farm data can be used to operationally represent decision-relevant tendencies under uncertainty, providing exploratory decision support applicable in practical farm settings.
This study examines the determinants influencing investors’ adoption of artificial intelligence (AI) as a decision-support tool in financial investments. Employing a quantitative research design, data were collected from 400 respondents who either currently utilize AI for investment activities or express an intention to do so. An online questionnaire served as the primary instrument for data gathering. Descriptive statistics—including frequency, percentage, mean, and standard deviation—were used to summarize participant characteristics, while inferential analyses employed an Independent-Sample t-test and structural equation modelling at a 0.05 significance level using software. The findings reveal that personal factors, such as demographic characteristics and investment experience, exert a significant influence on the decision to adopt AI tools, though certain variables exhibit similar effects across groups and others show no notable differences. Technological factors—including perceived usefulness, ease of use, and trust in AI systems—consistently demonstrate a significant positive relationship with the intention to employ AI for investment decision-making. No technological dimension was found to be insignificant, underscoring the critical role of technology readiness and system reliability. These results highlight the importance of both individual and technological considerations in shaping investor behaviour toward AI adoption. The study provides practical implications for financial technology developers, investment service providers, and policymakers seeking to enhance user confidence and promote broader utilization of AI in investment contexts.
Carpal Tunnel Syndrome (CTS) is a prevalent musculoskeletal condition that remains insufficiently understood by the public, particularly among younger adults who often underestimate their susceptibility. Prior studies have reported low public awareness of CTS symptoms and risk factors, contributing to delayed recognition and preventive action. To address this knowledge gap, this paper presents Wrisie, a vertical health chatbot developed using the Engati low-code framework to support CTS awareness, education, and basic symptom screening. Wrisie employs a knowledge-based conversational design that delivers structured CTS information through text, images, videos, and interactive exercises, enabling users to explore topics such as symptoms, risk factors, prevention, and treatment options. In addition, the chatbot includes a simple rule-based screening module that provides users with an indicative assessment of CTS symptom severity for awareness purposes only. A User Acceptance Test involving 30 participants was conducted to evaluate usability, clarity of information, interface design, and overall user satisfaction. The results indicate positive user perceptions of the chatbot’s usability and educational value, aligning with prior findings on the effectiveness of chatbots for health literacy and user engagement. The study demonstrates that a disease-specific, knowledge-based chatbot can serve as a practical tool for improving public health awareness and supporting early self-assessment, while complementing, not replacing, professional medical consultation.
We study ex-ante insurance design for systemic liquidity risk in an Eisenberg–Noe interbank network. Banks pay premiums proportional to external assets, and actuarial fairness requires that aggregate premium income equals expected aggregate shortfall losses. Within this framework, we characterize equilibrium premium determination as the fixed point where total premiums and expected losses coincide. We show that expected loss is a non-decreasing, piecewise linear function of the premium rate, with slope changes at default-regime transitions. For each fixed distressed set, we derive a closed-form marginal sensitivity based on the Leontief inverse of the distressed subnetwork, which clarifies how network topology governs premium incidence and contagion amplification. The analysis highlights that the equilibrium premium depends not only on shock magnitudes and aggregate capitalization, but also on who defaults jointly and how those institutions are interconnected. A numerical example demonstrates multiple equilibria, including a degenerate near-zero solution and a policy-relevant non-trivial solution, and illustrates how modest rewiring of interbank liabilities can shift the equilibrium premium upward by changing propagation channels. These results provide a transparent analytical basis for compulsory systemic liquidity insurance design.
Visual content plays a central role in online communication, where images are often accompanied by human interpretation and commentary. With the rapid development of Vision–Language Models (VLMs), artificial intelligence systems are increasingly capable of generating such interpretations. This paper examines whether current VLMs can produce art descriptions and interpretations comparable to those written by human students. We conduct a comparative study using a structured evaluation framework in which a Large Language Model (GPT-4o) acts as an automated judge assessing interpretations across seven dimensions, including symbolism, contextual awareness, originality, and coherence. Although the LLM-judge evaluates models higher than humans, experiments with GPT-4o, LLaVA1.5, Qwen2.5-VL, and Gemma3 reveal distinct differences. While humans map abstract concepts onto spatial reality and personal metaphor, advanced VLMs rely on detached, formal art critique terminology. Crucially, our variance analysis debunks the assumption that higher generation temperatures induce creativity; advanced models like GPT-4o exhibited zero or small variance, acting as unyielding, consistent critics. On the other hand, smaller models like LLaVA exhibited erratic, above-human variance due to abstract instability. These results highlight a critical design trajectory for multimodal AI: choosing between the rigid reliability of an automated “art critic” and the unpredictable, divergent brainstorming of an “idea-giver.”
Live-streaming shopping has become a highly interactive form of social commerce in which sustaining viewers’ participation is increasingly important. While prior studies have mainly focused on purchase intention and engagement, less attention has been given to how real-time interaction develops into attachment and stickiness. Drawing on the Stimulus–Organism–Response framework, this study integrates interaction ritual theory and social media affordances to examine how interaction rituals and platform affordances shape collective effervescence, attachment, and stickiness in live-streaming shopping. Data were collected from 249 users of Shopee, Facebook, Instagram, and YouTube Live in Taiwan, and analyzed using PLS-SEM. The results show that interaction rituals significantly enhance collective effervescence, which further strengthens emotional attachment. Emotional attachment positively influences platform attachment, and both forms of attachment significantly increase stickiness. In addition, metavoicing and guidance shopping significantly promote interaction rituals, while guidance shopping also strengthens platform attachment. This study contributes to the literature by explaining live-streaming stickiness as a socio-emotional and process-based outcome rather than a purely transactional response.
This study examines how existing explanatory approaches to the Gamer’s Dilemma can be applied to understand the divergence of ethics between the virtual and real worlds in Chinese Wuxia (武侠) games. The existing literature can be broadly grouped into three major approaches: moral intuition, game narrative, and social norms. However, these studies have focused primarily on Western games, and there remains insufficient examination of whether such approaches are equally applicable to Eastern genres such as Wuxia, a major game genre with broad influence across East and Southeast Asia. Wuxia is characterized by a relatively stable ethical structure centered on the principle of Yi (义). Shaped by the long-term interplay of historical, cultural, and social force, this normative framework has subsequently been carried over into and reproduced within Wuxia games. Taking the Chinese Wuxia game Where Winds Meet as a case, this study analyzes the moral judgments embedded in both its narrative design and gameplay system. The analysis indicates that both the game’s value orientation and its rule systems are consistently structured by Wuxia ethics, jointly shaping player conduct within this ethical framework. Among the three approaches, the social norms approach provides the most persuasive account, since Wuxia ethics function as a historically rooted and culturally shared normative system that delineates the boundaries of permissible action. The study thereby extends the Gamer’s Dilemma debate into a cultural domain that has received relatively little scholarly attention.
Sentiment analysis plays a critical role in disaster contexts, where social media provides large volumes of user-generated text. However, such text often contains noise, informal expressions, and irrelevant characters that can affect sentiment classification performance. This study proposes a pipeline called CLEAN that leverages a multilingual pretrained language model to classify sentiment in disaster-related social media data, while employing large language models (LLMs) for automated text cleaning and normalization. Experiments on disaster-related social media text show that sentiment classification on cleaned text consistently outperforms classification on raw text. These findings suggest that LLM-based preprocessing can enhance the reliability of sentiment analysis for disaster response applications.
Optical Character Recognition (OCR) for real-world receipt images is affected not only by clean-image recognition performance but also by robustness under image degradation and language-specific error behavior. In this work, we present an empirical study of five OCR engines on Vietnamese receipt images using controlled perturbations, including blur, noise, rotation, and JPEG compression. We evaluate the systems with Character Error Rate (CER), Word Error Rate (WER), and Exact Match Rate under a unified text normalization protocol, and further analyze error patterns through edit operations, character confusion statistics, and Vietnamese-specific diacritic error categories. Our results show that the best-performing OCR engine on clean images is not necessarily the most robust under degraded conditions: VietOCR achieves the strongest clean and overall accuracy, while PaddleOCR is the most stable engine under degradation. Blur is the most harmful perturbation across engines, and different systems exhibit distinct failure mechanisms, especially in handling Vietnamese diacritics and base-character variants. These findings suggest that OCR evaluation for Vietnamese documents should go beyond aggregate CER and separately consider clean accuracy, robustness, and linguistic fidelity.
This study presents a preliminary computational analysis of viewer engagement within the burgeoning domain of digital agricultural communication. While traditional metrics focus on interaction volume, this research explores the structural properties of online viewer communities. Using a dataset of 772 YouTube videos, we categorized content into four themes: agricultural product, regional scenery, production method, and rural life. By integrating VADER sentiment analysis and Social Network Analysis (SNA), we identify significant variations in how different content stimuli shape social network topology. Initial results suggest that high-frequency interaction (Engagement Rate) does not inherently align with network stability (Cohesion). This paper outlines the methodological workflow and early structural observations.
While automated brain tumor detection can greatly improve patient outcomes, many existing models remain ’black boxes’ that lack the transparency and reliability needed for real-world clinical use. To address this, we propose a unified MRI analysis framework that simultaneously performs slice-level classification and 2D tumor segmentation using the BraTS 2020 dataset. Instead of merely providing a binary diagnosis, our classification branch–built on EfficientNetV2-B3–incorporates self-attention and evidential prediction. This allows the system to not only flag tumors but also explain its focus and quantify its own predictive uncertainty. For precise localization, the segmentation branch utilizes a DeepLabV3+ architecture with a ResNet101V2 encoder, leveraging ASPP and feature fusion to capture complex multi-scale context. In our evaluations, the classification model achieved 89.67
In recent years, the development of the Internet has led businesses to increasingly focus on collecting online public reviews for text mining, with sentiment analysis emerging as a prominent application in this field. However, obtaining a balanced dataset remains a significant challenge. Data imbalance can introduce significant biases in the model’s training and prediction processes, decreasing the predictive performance of minority classes. To address this issue and mitigate the effects of data imbalance, this study employs generative adversarial network (GAN) models as a data augmentation technique to generate sufficient samples to correct the imbalance in the dataset. Furthermore, it compares the sentiment prediction results before and after the application of the data augmentation model. The experimental results indicate that the additional sample data generated through the data augmentation model significantly improves the accuracy of the sentiment analysis model, thereby optimizing and enhancing the machine learning model.
Spam reviews pose a serious threat to Vietnamese E-commerce platforms, as malicious actors increasingly exploit review systems to manipulate product rankings, mislead consumers, and erode marketplace trust. This problem is compounded by the prevalence of Code-Mixed writing, where Vietnamese text is combined with English keywords, brand specific terminology, and obfuscated expressions that challenge conventional spam detection methods. Unlike prior work that relies on either monolingual or multilingual encoders in isolation, we propose a Code-Mixed Transformer Fusion model that processes each review in parallel using PhoBERT and XLM-R to jointly capture Vietnamese specific linguistic patterns and multilingual spam related signals. The proposed architecture integrates local attention pooling, global sentence level representations, and a gated fusion mechanism to construct a unified review embedding for robust spam classification. To further enhance detection performance under class imbalance, an optimal decision threshold is selected through F1 score maximization on validation data. Experimental evaluation on Vietnamese E-commerce review datasets demonstrates that the proposed model consistently outperforms strong baseline approaches, achieving up to 94.47
