With the advent of generative artificial intelligence (AI), which has drastically reduced the cost of content generation, the design of visual and textual advertising for online tourism products has gained new significance. However, how to design effective multimodal advertising remains underexplored. Prior research has shown that such advertising influences consumer decisions, yet few studies have extracted hierarchical visual cues in an interpretable manner and have quantified their predictive value for promotional outcomes. We propose a multimodal prediction framework based on a multimodal large language model (MLLM) for tourism advertising effectiveness. The framework extracts visual structural cues with LayoutXLM and abstract visual content cues with an MLLM, then fuses them via a dual-gated module to predict advertising effectiveness, providing quantitative support for evaluation and deployment. We conducted experiments on 8640 products collected from Ctrip over a period of five months. The results indicated that fusing visual structural features and content features enabled the model to achieve the best performance. Compared with the baseline, the MSE decreased by 0.2396. The RMSE decreased by 0.1280. The R2 increased by 0.1541. This study provides merchants with practical guidance on advertising design and offers online travel platforms data-driven strategies for advertising placement.

Streamers’ discourse plays a key role in shaping purchasing decisions in live streaming e-commerce, especially in knowledge-sharing formats where product promotion is combined with information delivery. Previous studies have shown that streamers’ discourse can influence product sales, with few studies systematically extracting semantic features across different dimensions and quantifying their impact on sales prediction performance. The main contribution of our research is the design of a predictive framework for sales in knowledge-sharing live streaming. The framework integrates social support theory with fine-tuned large language models (LLMs) to systematically extract multi-dimensional semantic cues from streamers’ discourse for sales prediction. We collected data from 80 live streams across 35 Douyin rooms over two months for our experiments. In the social support classification experiment, the fine-tuned Ernie-SFT model outperformed the best baseline LLM, with improvements of 11.12% in accuracy, 11.87% in weighted F1-score, and 7.83% in macro F1-score. In the sales prediction experiments, we validated the proposed framework using four mainstream classifiers and observed consistent performance gains. The best-performing classifier achieved improvements of 12.53% in accuracy, 10.83% in weighted F1-score, and 4.24% in macro F1-score. These findings highlight the strong predictive value of social support features embedded in streamers’ discourse, offering actionable insights for streamers and enabling data-driven optimization strategies for platforms.

While large language models (LLMs) have revolutionized writing assistance, their widespread misuse poses significant threats to scholarly integrity and authorship originality. Existing detection methods, often relying on single-dimensional features and document-level binary classification, lack the granularity to identify localized LLM involvement and are vulnerable to sophisticated evasion tactics like paraphrasing. To address these limitations, we propose a novel Multi-level Fine-grained Detection (MFD) framework, which integrates low-level statistical, high-level semantic, and deep-level linguistic features to accurately localize LLM-generated content at the sentence level. Specifically, we employ a robustness-aware contrastive learning mechanism trained on adversarially rewritten texts to capture invariant high-level semantic features. Furthermore, we leverage an advanced LLM to extract deep linguistic nuances and contextual patterns that traditional methods overlook. Extensive experiments on public datasets demonstrate that MFD outperforms state-of-the-art methods, achieving an MAE of 0.1347 and an Accuracy of 0.8856. This work provides educational institutions with a precise and robust mechanism to mitigate the risks of compromised authorship, ensuring strict adherence to academic standards.

Text-to-SQL technology has evolved rapidly, with diverse academic methods achieving impressive results. However, deploying these techniques in real-world systems remains challenging due to limited integration tools. Despite these advances, we introduce Squrve, a unified, modular, and extensive Text-to-SQL framework designed to bring together research advances and real-world applications. Squrve first establishes a universal execution paradigm that standardizes invocation interfaces, then proposes a multi-actor collaboration mechanism based on seven abstracted effective atomic actor components. Experiments on widely adopted benchmarks demonstrate that the collaborative workflows consistently outperform the original individual methods, thereby opening up a new effective avenue for tackling complex real-world queries. The codes are available at https://github.com/Satissss/Squrve.

Text-to-SQL systems aim to enable users to query structured data using natural language, and recent advances in large language models have significantly improved their performance on general-domain benchmarks. However, existing approaches remain largely domain-agnostic and struggle to adapt to financial applications, where the diversity of SQL dialects across database management systems and the complexity of domain-specific semantics frequently lead to substantial degradation in both syntactic correctness and logical consistency. To address these limitations, we propose a financial-domain-oriented Text-to-SQL framework that jointly enhances SQL generation from two complementary perspectives. First, we introduce a syntax-aware mechanism that mitigates SQL dialect confusion and improves the robustness of code generation across heterogeneous financial databases. Second, we incorporate a semantic alignment module that captures implicit financial business rules and domain terminology, enabling the model to produce SQL that better reflects users' intent. Experimental results on a real-world financial dataset demonstrate that our method substantially reduces both syntax and logic errors, highlighting the necessity of domain-specific adaptation for reliable Text-to-SQL deployment in specialized environments.

The proliferation and rapid spread of fake news on social media pose a significant threat to society, underscoring the urgent need for effective early detection methods. This paper introduces multimodal adversarial transfer learning (MATRAL), a novel approach designed for early fake news detection. MATRAL integrates multimodal learning with adversarial transfer learning. Through effective multimodal learning, MATRAL can form a comprehensive representation of news items on social media, including text, images, and publisher information. The adversarial transfer learning component enables MATRAL to dynamically adapt its knowledge to new domains, ensuring the approach’s ongoing relevance against the evolving fake news generation tactics. Using the MediaEval 15–16 data sets to simulate the early fake news detection scenario, we conduct extensive experiments to evaluate MATRAL’s performance against state-of-the-art methods in multimodal fake news detection, machine learning methods, and industrial practices. The experimental results conclusively demonstrate MATRAL’s superiority across various widely adopted metrics, showcasing its proficiency in early stage fake news detection. To further elucidate the contributions of various model components, a series of ablation studies are conducted. Furthermore, MATRAL’s interpretability and robustness are substantiated through additional experimental analyses. Our work introduces a novel and robust solution to the pressing challenge of multimodal fake news detection on social media, offering a significant contribution to the research and practice of responsible artificial intelligence. History: This paper has been accepted by Kaushik Dutta for the Special Issue on the Responsible AI and Data Science for Social Good. Funding: This research was partially supported by the National Natural Science Foundation of China [Grants 72495123, 72101007, and 72201061]. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information ( https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2023.0514 ) as well as from the IJOC GitHub software repository ( https://github.com/INFORMSJoC/2023.0514 ). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/ .

Insufficient sales prediction easily leads to the untimely dispatch of supply and causes inventory problems, resulting in a loss of profits for merchants. Previous sales prediction research was mainly based on traditional e-commerce platforms that cannot be directly applied to live streaming e-commerce, which adds many important elements. The main contribution of our research is the design of a multimodal analytics framework for product sales prediction in live streaming e-commerce. In the proposed framework, we explore the influence of anchor reputation on product sales and innovatively consider both historical and real-time reputation signals. In addition, to better extract valuable information for real-time signals, we propose an A-tiFSR model to extract features from product text and images and design a fine-grained analysis method to mine danmaku data. We conduct our experiments based on a real-world dataset collected from Douyin live streaming. The experimental results demonstrate the effectiveness of the constructed multimodal reputation signals of anchors on product sales prediction. The management implication of our research is that merchants and anchors should pay more attention to their multidimensional reputation. The proposed sales prediction framework can help merchants predict sales more accurately and optimize inventory policy and marketing strategies. This study thus supports the more effective development of live streaming e-commerce markets.

Graph contrastive clustering (GCC) has achieved numerous advantageous results due to the information mining capability of self-supervised learning. Multi-view attribute graph clustering, as a means of addressing complex attribute graph data with multi-view relations or features, has gained significant attention in recent years. However, only a few GCC methods are applicable to multi-view graph data, limiting the full utilization of the rich information contained in such data. Furthermore, most GCC methods rely on random augmentation strategies, leading to structural unfairness in the augmentation process across views. To address these issues, we propose multi-view fair-augmentation contrastive graph clustering with reliable pseudo-labels (MFCGC), which combines contrastive learning and fully utilizes the node-level and cluster-level information within each view of the multi-view data, ensuring the consistency of information across different views. Moreover, we designed a carefully crafted node-degree-based augmentation method called fair augmentation, which preserves the partial topology structure of multi-view data. Finally, we propose a reliable pseudo-label selection mechanism that integrates reliable pseudo-labels from multi-view graph to improve the quality of sample pairs. Our proposed MFCGC has been extensively evaluated on real-world datasets, demonstrating its superiority over state-of-the-art algorithms in two categories of multi-view graph data and attesting to its significant effectiveness.

As a new shopping paradigm, live streaming e-commerce has undergone rapid development in recent years. However, a higher return rate than that of traditional e-commerce not only affects the consumption experience of users but also is detrimental to the development of the entire industry. Previous studies on product returns have mainly focused on traditional e-commerce, with few studies considering signals in unstructured data formats. The main contribution of our research is the design of a multimodal analytics framework to predict product returns in live streaming e-commerce settings. Multimodal features (e.g., visual, textual, and acoustic) and their correlations extracted from different data sources (product descriptions and live streaming records) are considered in our prediction model. In addition, to better study the impact of anchor speech during live streaming, we propose an acoustic feature extraction model called “OpenTransformer,” which extracts acoustic features of anchor audio and outperforms baselines. We have conducted experiments based on a real-world dataset collected from Taobao live streaming. The experimental results demonstrate the effectiveness of using multimodal signals from products and anchors for product return prediction. The managerial implications of our research are that live streaming platforms and potential consumers can prejudge which sold products will have high return rates in a timely fashion, and based on these predictions, merchants and anchors can better design their product descriptions and speech behaviors. This study thus supports more effective development of live streaming e-commerce markets.

Viewer gifting is an important business mode in live streaming industry, which closely relates to the income of the platforms and streamers. Previous studies on gifting prediction are often limited to cross-section data and consider the problem from the macro perspective of the whole live streaming. However, the multimodal information and the time accumulation effect of live streaming content on viewer gifting behavior are ignored. In this paper, we put forward a multimodal time-series method (MTM) for predicting real-time gifting. The core module of the method is the multimodal time-series analysis (MTA), which targets at effectively fusing multimodal information. Specifically, the proposed orthogonal projection (OP) model can promote cross-modal information interaction without introducing additional parameters. To achieve the interaction of multi-modal information at the same level, we also design a stackable joint representation layer, which makes each target modality's representation (visual, acoustic and textual modality) can benefit from all the other modalities. The residual connections are introduced as well to ensure the integration of low-level and high-level information. On our dataset, our model shows improved performance compared to other advanced models by at least 8% on F1. Meanwhile, the MTA is able to meet the real-time requirements of the live streaming setting, and has demonstrated its robustness and transferability in other tasks. Our research may offer some insights about how to efficiently fuse multimodal information, and contribute to the research on viewer gifting behavior prediction in the live streaming context.

Initial coin offerings (ICOs) provide an early-stage financing method for blockchain-based ventures. During the ICO process, whitepapers are important not only as promotional material through which ventures can demonstrate the technical and financial project details but also as references for investors. Persuasion theory and the related literature suggest that the presentation and order of information have a significant impact on the attitude of the audience. Therefore, in addition to projects’ metadata features, we construct a document analysis deep regression model (DADRM) to innovatively extract deep text and layout features from whitepapers. Based on a real-life dataset, we conduct a comparative study to assess the effectiveness of the proposed framework in predicting ICO success in terms of the funding amount. The empirical results show that our model that both extracts text content and retains the original 2D structure of the document can significantly reduce prediction error. Based on our proposed model, both ICO platforms and investors can prejudge the funding amount of cryptocurrency projects and mitigate information asymmetry. Additionally, this study demonstrates that both what is written in the business document and how the document is presented affect investor decisions.

Initial coin offering (ICO) is a new financing method that has been widely used in cryptocurrency projects. However, it has been reported that nearly 30% of cryptocurrency projects fail during ICO, indicating an important gap in research and an opportunity for more advanced research on ICO project assessment. This study reveals that previous studies primarily used project-related factors to predict ICO success while neglecting social factors such as team information and expert evaluation. Inspired by the knowledge-based theory (KBT) of the firm, we set out to examine the impact of heterogeneous team knowledge and expert evaluation on ICO success. One primary contribution of this study is the design of novel knowledge measures based on KBT. In addition, we propose a deep-learning model – an attention-based bidirectional recurrent neural network (A-BiRNN) – to automatically extract features from online comments. We validate the proposed model on a real-world dataset, and experiments show that the accuracy of the proposed prediction model outperforms those of existing models by more than 6%, highlighting the effectiveness of the proposed approach in predicting ICO success. This study's results provide useful ideas for both investors and ICO platforms to assess the quality of cryptocurrency projects, thus improving information symmetry in ICO markets. Also, this study demonstrates the value of applying KBT in assessing firm performance in ICO markets. The generalized value of the proposed approach should be tested in more business contexts, such as crowdfunding and peer-to-peer (P2P) lending.

Danmaku is an emerging comment design for videos that allows real-time, interactive comments from viewers. Danmaku increases viewers’ interaction with other viewers and streamers, thereby raising viewers’ loyalty and sense of belonging. Sending Danmaku comments demonstrates a higher degree of viewer involvement than traditional static comments below the videos. Therefore, it is necessary and meaningful to learn about viewers’ preferences by observing their behavior, as this may benefit the platform as well as the streamers. However, research on how the multimodal environment affects viewers’ behavior in sending Danmaku comments is quite limited. To fill this gap, we propose a new dataset and a deep neural network integrating multimodal information to predict whether viewers will send Danmaku comments (Deep Multimodal network for Danmaku Forecasting, DMDF) in order to evaluate the impact of the interaction of textual features, audio features and visual features on the behavior of viewers sending Danmaku comments. A series of experimental results based on a real dataset of 249657 samples from Bilibili (a leading Chinese video streaming Website) demonstrate the effectiveness of the proposed DMDF and the helpfulness of all modalities, especially visual and acoustic features, in behavior forecasting. DMDF with the multimodal squeeze-and-excitation (MSE) module we proposed achieves 90.14% on accuracy and 83.60% on F1-score, and it reveals the extent to which a user-generated video can influence viewers to send Danmaku comments, which helps predict viewers’ online viewing behavior. Furthermore, our model contributes to the current work on the video understanding task.

PurposeMicro-video platforms have gained attention in recent years and have also become an important new channel for merchants to advertise their products. Since little research has studied micro-video advertising, this paper aims to fill the research gap by exploring the determinants of micro-video advertising clicks. We form a micro-video advertising click prediction model and demonstrate the effectiveness of the multimodal information extracted from the advertisement producers, commodities being sold and micro-video contents in the prediction task.Design/methodology/approachA multimodal analysis framework was conducted based on real-world micro-video advertisement datasets. To better capture the relations between different modalities, we adopt a cooperative learning model to predict the advertising clicks.FindingsThe experimental results show that the features extracted from different data sources can improve the prediction performance. Furthermore, the combination of different modal features (visual, acoustic, textual and numerical) is also worth studying. Compared to classical baseline models, the proposed cooperative learning model significantly outperforms the prediction results, which demonstrates that the relations between modalities are also important in advertising micro-video generation.Originality/valueTo the best of our knowledge, this is the first study analysing micro-video advertising effects. With the help of our advertising click prediction model, advertisement producers (merchants or their partners) can benefit from generating more effective micro-video advertisements. Furthermore, micro-video platforms can apply our prediction results to optimise their advertisement allocation algorithm and better manage network traffic. This research can be of great help for more effective development of the micro-video advertisement industry.

Online healthcare communities (OHCs) have become producers of medical information. Solving the issue of how to effectively reuse such a large amount of medical data and discover its potential value is of the utmost importance for alleviating the shortage of medical resources. Online consultation has received widespread attention and population since its first appearance in 1999, and as a result, many diagnostic multi-turn questions and answers (Q&A) documents have become available. This type of document is formed by multiple rounds of patient questions and doctors’ diagnostic answers and contains massive medical knowledge and doctors’ diagnostic experience. Few studies concentrate on the modeling and recommendation of this type of document, yet making these documents convenient for reuse reduces the cost of medical consultation for patients and saves time addressing common diseases for doctors. In this paper, we focus on the modeling and understanding of diagnostic multi-turn Q&A records and propose a deep-learning recommendation framework based on patient medical information needs, the contents of Q&A records and doctor background information. With the evaluation based on a real dataset that contains pediatric consultation dialogues fetched from DingXiangYuan, a famous online consultation application in China, we found that the proposed model achieved a good performance on the recommendation of diagnostic multi-turn Q&A records and outperformed baseline models. In addition, we discussed a potential application scenario of the recommendation model, suggesting that the proposed model can promote the reduction of patient costs and doctors’ work pressure in countries or regions with insufficient medical resources.

A domain-dependent customer requirements mining framework to facilitate mobile app quality upgrades is proposed in this paper. We develop a new ranking model to rank the importance of different customer requirements by considering both the rating data and review data. We prove the effectiveness in terms of product quality improvements based on 265 version update cases for 15 popular mobile apps. As there is little research regarding identifying the business value of customer requirements mining, this study can be highly beneficial to the further development of research concerning the business value of adopting online customer requirements for product improvements.

With the rapid development of e-commerce, a new type of secondhand e-commerce website has appeared in recent years. Any user can have his or her own shop and list superfluous items for sale online without much supervision. These secondhand e-commerce platforms maximize the economic value of secondhand markets online, but buyers risk conducting unpleasant transactions with low-reputation sellers. The main contribution of our research is the design of a text analytics framework to assess secondhand sellers' reputation. In addition, we develop a new aspect-extraction method that combines the results of domain ontology and topic modeling to extract topical features from product descriptions. We conduct our experiments based on a real-word dataset crawled from XianYu. The experimental results reveal that our ontology-based topic model method outperforms a traditional topic model method. Furthermore, the proposed framework performs well in different item categories. The managerial implication of our research is that potential buyers can prejudge the reputation of secondhand sellers when making purchase decisions. The results can support a more effective development of online secondhand markets.

Merchants, as well as customers, have noticed the importance of online product reviews and numeric ratings in electronic commerce websites. It is valuable if merchants can discover some potential customer value from the sheer volume of data. This paper contributes a semantic text analytics approach that can dig out the customers' most basic concerns about their online purchase choices. More specifically, based on the hypothesis that the product reviews and overall ratings estimated by same person in a tiny time interval have a great relevance, we dexterously utilize this relevance to realize the embedded customer value. In the proposed method, take the single lens reflex camera for example, an innovative aspect extraction method that comprehensively considers the product ontology and results of the topic modeling method latent Dirichlet allocation is applied. As a result, 8 specific aspects are identified from the experimental results. For each aspect, a self-contained review feature corpus is created as an extension of some seed terms. After aspect-based sentence segmentation and context-sensitive sentiments preprocessing, aspect-oriented sentiment analysis is applied. Multiple regression analysis is then used as a statistical measure to discover determinant aspects of overall ratings. The results reveal that cost performance, image quality and product integrity are the three most influential aspects. The practical implication of our research is that merchants can efficiently modify their products, to satisfy more customers and also boost sales performance.

As crop diseases bring huge losses every year in both developed and developing countries, determining how to precisely predict crop disease severity to facilitate agricultural emergency management is really a worldwide problem. Previous studies have introduced machine learning (ML) techniques into crop disease prediction and achieved better experimental results. However, the architectures of these ML models are unsuitable to model time series data. Moreover, the dependences among observations over time and across space have not been taken into account in model construction. By applying data-mining techniques to dynamic spatial panels of remote sensing data and considering features of bioclimatic, topographic and soil conditions as a supplement, we propose a novel crop disease prediction framework for agricultural emergency management based on ensemble learning techniques and spatio-temporal recurrent neural network (STRNN) which is an extension of recurrent neural network (RNN) in time and space. Empirical experiments are conducted on a specific dataset which is built based on reported cases of wheat yellow rust outbreaks in the Longnan city. Experimental results indicate that our proposed method outperforms all baseline models in crop disease severity prediction. The managerial implication of our work is that by applying the proposed methodology, some preparedness measures can be implemented in advance to prevent or mitigate the possible disasters according to predicted results. Notable economic and ecological benefits can be achieved by optimizing the frequency and timing of application of fungicide, pesticides and other preventative measures.