Organizations are beginning to deploy causal machine learning (causal ML) models for estimating average treatment effects and treatment effect differences for various subgroups, as well as identifying causal mechanisms driving treatment effects. For instance, in domains such as health care and marketing, causal ML is being used to identify who is most likely to benefit from or be receptive to treatments or interventions, including medical treatments and marketing campaigns. However, while causal ML models can produce highly accurate results when causal assumptions are met, the outputs can be confusing to work with. For instance, it is easy to conflate risk scores with conditional average treatment effects (CATEs), as risk scores from predictive models identify who is most likely to experience an outcome while CATEs from causal models identify who is most likely to benefit (or not benefit) from a treatment. A practitioner reviewing causal ML results may not fully understand such nuances and, thus, may not effectively convert relevant causal ML outputs into practice. Thus, a particular challenge relevant to causal ML, and extensible to any complex set of digital artifact outputs, is that effective practice transformation requires effective boundary spanning. To address this challenge, we propose a framework that conceptualizes large language models (LLMs) as dynamic boundary spanning objects. The framework leverages different forms of human-LLM collaborations throughout the causal ML boundary spanning process, toward the goal of successful practice transformation. We further provide an example of how the framework can be applied to a case of complex causal ML results. We ultimately seek to extend boundary spanning theory and provide a pathway for translating causal ML results across syntactic, semantic, and pragmatic boundaries.
Recommendation algorithms are crucial to social media platforms, determining the order of information users see and aiding in discovering news. However, they can inadvertently cause harm through algorithmic privilege, excessive personalization, filter bubbles, and misinformation recommendation and propagation. Spreading misinformation through online social networks like Facebook, X (formerly Twitter), TikTok, and Instagram is commonplace in today’s milieu. Similar such cases were during the COVID-19 pandemic and the US Presidential election where misleading information spread widely. Such issues highlight the need for transparency and oversight in recommendation algorithms design. While some research has been conducted on mitigating misinformation within recommendation algorithms, the focus has primarily been on intervention strategies, which have been observed to cause a backfire effect. Likewise, the current literature only focuses on addressing misinformation at the time of recommendation, particularly without considering dissemination which is very concerning. Our study focuses on strategies to generate recommendations by modeling trustworthy social network and/or user neighbors aggregation while designing user-item latent factors to mitigate misinformation recommendation and dissemination. We evaluated various trustworthy user modeling designs on two Twitter datasets to assess the misinformation in top-K recommendation lists and simulated the propagation of recommended misinformation within the user network to also analyze misinformation spread. The research findings indicate that trustworthy social network and/or user neighbors modeling helps to mitigate misinformation spread at both at time of recommendation and dissemination.
The advent of generative artificial intelligence (AI) has heightened the proliferation of fake news. A key challenge is the limited real-world data to investigate the societal impact of fake news produced by generative AI. In this article, we examine stock market reactions to financial news articles that exhibit stylometric similarity to human-crafted and AI-crafted fake financial news. Grounded in language expectancy theory, we employ a style-based transfer learning model, pre-trained to recognizing deceptive language employed in various types of fake news intricacies. We then apply this model to a comprehensive dataset of financial news, assigning a “veracity style score” to each article. This score quantifies the extent to which an article’s language and stylistic features align with patterns typically observed in human-crafted or AI-crafted fake news, indicating the risk of deception. By analyzing these scores of financial news articles on S&P 500 stocks, we find differential market reactions. Deceptive language in news articles stylometrically similar to human-crafted fake news is negatively associated with abnormal trading volume and absolute abnormal returns, whereas deceptive language in news articles stylometrically similar to AI-crafted fake news is positively associated with both metrics, highlighting the risk of market inefficiency in identifying AI-crafted deceptions.
The rapid growth of textual data from social networks and online reviews has amplified the importance of research on cross-domain fine-grained sentiment classification. To address challenges such as data sparsity and domain discrepancies in current multi-source cross-domain sentiment classification (MSDFSC) methods, this study introduces a novel approach. This study proposes a multi-source cross-domain fine-grained sentiment classification method based on domain combination and voting ensemble from a meta-learning perspective (MLDCV-FGSC), the framework involves three stages. First, during data partitioning, we employ meta-learning by designating one of four domains as the target domain and the others as source domains. The data for each source domain are divided into support sets and query sets. The meta-training dataset is then constructed by combining the support sets from two source domains as the support set for a meta-task, while the query set from another source domain serves as the query set for the meta-task, thereby creating multiple meta-learning tasks. Secondly, during the meta-training stage, we train a base classifier for each meta-task employing a BERT model to encode both the text and the aspect terms. A multi-head attention mechanism is further incorporated to enhance the interaction modeling between them. The classifier parameters are updated through backpropagation to progressively improve its performance in cross-domain sentiment classification. Finally, in the meta-testing stage, we fine-tune the base classifiers to adapt to the target domain. The final prediction is obtained through an ensemble voting mechanism that integrates the output of the fine-tuned classifiers, ensuring robust generalization to the target domain. Our experimental validation, using datasets from various domains and various sentiment classification tasks, demonstrates that the proposed method achieves significant performance improvements in cross-domain fine-grained sentiment classification. Furthermore, multi-modal experiments conducted on other datasets further validate its effectiveness. These results underscore the efficiency and substantial potential of our approach in overcoming cross-domain sentiment classification challenges.
Given the profound impact of diet on cardiovascular health, appropriate dietary intake in home cooking is widely acknowledged as being instrumental in preventing cardiovascular diseases (CVDs). However, contemporary dietary recommendation systems for patients with CVDs are limited by population-level guidelines, static advice, as well as a lack of consideration for contextual conditions and individual preferences. We hence subscribe to the Theory of Planned Behavior (TPB) to design a novel recipe recommender system. Our proposed recipe recommender system takes into account users’ personal attitudes, subjective norms, and perceived behavioural control as focal considerations in recipe recommendations. Particularly, our recipe recommendation system assimilates personalized considerations—including health conditions, taste preferences, and available ingredients in the user's refrigerator—with social elements in the likes of cooking frequency, recipe viewing intensity, and user comments to bolster the acceptability of recommended recipes. Additionally, by incorporating user interactions in the likes of adding and liking recipes, our proposed recipe recommender system streamlines recipe discovery and strengthens users’ perceived behavioural control to maintain dietary choices. This integration aims to alleviate the practical challenges associated with adopting recommended recipes, rendering it much easier for users to adhere to dietary guidelines. Based on empirical validation, we not only deliver detailed insights into the implementation of a recipe recommendation system, but we also validate its utility in enhancing user engagement and satisfaction.
