9월 5주차

Machine-learning-based artificial intelligence (AI) systems often perform well in controlled settings but fail when deployed in the real world. These failures stem from a fundamental problem: ML/AI models optimize data-model fit by increasing statistical performance on training data, whereas real-world deployment demands a new paradigm in ML/AI: model-world fit that preserves domain semantics and contextual understanding. We propose a Machine Learning Model-World Fit framework that provides both a conceptual structure for understanding how domain knowledge shapes ML/AI systems and a research roadmap for investigating these relationships. The framework identifies how domain knowledge can be integrated into ML/AI development process via external knowledge representations, such as data, value, process models or domain ontologies. Incorporating external knowledge representations can help overcome three types of knowledge deficiencies in AI: know-what, know-how, and know-why. To illustrate the framework, we apply it to a child welfare placement case, a domain where contextual understanding is critical and statistical accuracy is an insufficient evaluation criterion. Our insights lead to several directions for scholars working at the intersection of machine learning, design science, and knowledge representation.

Online labor markets play a vital role in connecting freelancers and employers worldwide. While most existing methods focus on helping employers find the best candidates, relatively little research addresses freelancer-centric approaches. In this study, we propose a freelancer-centric skill advisor (FESA) that combines a competitiveness evaluator and an optimization model. The competitiveness evaluator integrates a state-of-the-art deep learning recommender system with contrastive learning to calculate matching scores between freelancers and their desired tasks, while the optimization model prescribes skills a freelancer should acquire to improve their competitiveness for those tasks while making the learning experience manageable. FESA also supports two reskilling strategies: exploration and exploitation. We evaluate FESA against alternative recommendation models using a large-scale, real-world dataset collected from a leading online labor market. We also conduct a controlled experiment with human participants to examine how real-world users respond to FESA recommendations and whether disclosing a freelancer’s ranking, a key output of FESA, affects reskilling motivation. In contrast to prior findings from traditional organizational settings, where ranking information generally encourages greater effort, we find that disclosing a freelancer’s low ranking significantly reduces motivation to reskill. This demotivating effect of ranking disclosure is more pronounced under the exploration reskilling strategy, which requires freelancers to acquire more challenging and unfamiliar skills, than under the exploitation reskilling strategy where freelancers are advised to acquire skills that are more closely aligned with their existing expertise. Our research highlights that designing a freelancer-centric recommendation system requires not only algorithmic innovation but also careful attention of how system design influences user motivation and engagement.

Children increasingly consume video content through platforms such as YouTube and YouTube Kids, yet the content-level drivers of their engagement remain poorly understood. Unlike adults, children rely less on social and participatory cues in digital media, suggesting that their attention may be shaped more directly by perceptual features. In this research, we investigate visual motion, quantified using optical flow, as a perceptual feature that shapes children’s engagement with digital video. Across three studies, we evaluate its influence using observational, field experimental, and behavioral data. In Study 1, we analyze more than 32,000 videos from leading children’s channels on YouTube and find that higher visual motion significantly predicts video viewership after accounting for a broad set of visual, audio, and affective characteristics. In Study 2, we create educational videos with experimentally manipulated background motion and deploy randomized advertising campaigns using YouTube’s A/B testing infrastructure. Videos with higher motion generate significantly higher view-rates and views, replicating the observational findings in a field setting. In Study 3, a lab-based eye-tracking experiment with children shows that visual motion increases screen engagement primarily by reducing the duration of long off-screen attentional lapses rather than reducing the frequency of attention breaks. Additional analyses show that motion increases attention to background regions without reducing attention to instructional content, a pattern more consistent with screen-level re-engagement than with the displacement of instructional attention. These findings identify visual motion as a perceptual mechanism that reduces prolonged disengagement by accelerating attentional re-orientation, with implications for the design and governance of child-directed media.

Geofencing, with the advent of location-based services, enables an advertiser to set a virtual fencing zone around its establishment and inform consumers within the zone to visit it. Yet, when to adopt and how to optimize geofencing remains unclear, despite its popularity across industries. We investigate the optimal geofencing using a game-theoretic model, highlighting the importance of initial consumer awareness in geofencing configurations. We consider an advertiser whose consumers are heterogeneous in location (i.e., local or remote) and awareness level. We first compare geofencing and mass advertising under location-independent awareness (i.e., the ratio of informed and uninformed consumers is consistent in both locations). We show that geofencing outperforms mass advertising as the advertiser focuses on the local market. Two benefits support the advertiser’s localization, namely, profit-margin enhancement and cost savings. We then obtain the optimal fencing radius and price under various market conditions with respect to product valuation and initial awareness. We further extend our location-independent awareness model to a location-dependent awareness, wherein the initial awareness level is negatively associated with the distance to the advertiser. The findings showcase the significant impact of the awareness distribution on the applicability and profitability of geofencing, which is underexplored in the literature. Last, to improve the efficiency of geofencing, we propose flexible geofencing that encourages the advertiser to flexibly choose the point of interest, not always staying with its store location. Our analytics detail the superiority of flexible geofencing over the typical proximity-based practice in the scenario of location-dependent awareness. Such a flexible strategy simultaneously improves the advertiser’s profits, consumer surplus, and social welfare. In sum, we voyage in the ocean of advertising, sail our research to the uncharted treasure island of geofencing, and map a course to take new adventures for future research and practices.