I build AI systems,
then measure what
they change.

Purpose. This study examines the causal impact of YouTube Shorts' ad-revenue sharing policy on creator production strategies in the short-form video creator economy.

Design. Using 36,507 videos from 342 top global channels, we employ a regression discontinuity in time design. Computer vision techniques measure four attributes of each creator's production strategy: output volume, video length, editing pace, and content novelty.

Findings. The policy drove an immediate reduction in video length without expanding output volume, consistent with substitution toward cheaper per-unit configurations rather than output scaling. Entertainment creators shortened their videos most sharply, while informative creators slowed their editing pace to preserve comprehension. Long-type creators pursued per-unit cost reduction, while short-type creators invested in content novelty.

Implications. The quantity-expansion result from low-marginal-cost user-generated content settings does not extend to high-marginal-cost production, where incentives reallocate effort within existing output rather than expand it. A uniform monetization policy does not produce uniform outcomes.

KrAIS Summer Workshop 2023, Seoul.

Online dating platforms increasingly mediate human connection, yet the mechanisms by which users evaluate subjective traits like appearance and personality remain underexplored. To investigate these preference dynamics, this study applies AI-driven methods to a dataset of 506,014 interactions from 41,441 users on a major heterosexual dating platform in South Korea. By employing computer vision and large language models to quantify facial and personality similarity, our mixed-effects analysis reveals a significant gender asymmetry: women prefer facial similarity (homophily), whereas men exhibit a preference for facial dissimilarity (heterophily). Furthermore, personality preferences are found to be context-dependent; a partner's socioeconomic status moderates the demand for similarity, amplifying the preference for women while attenuating it for men in upward evaluation contexts. These findings offer actionable implications for the design of matching systems, suggesting that preferences in digitally mediated environments function as adaptive strategies shaped by gender and contextual cues.

Online dating has become a social infrastructure, with over one-third of marriages originating from digital platforms. Despite its significance, our understanding of how users form impressions and make matching decisions remains limited. This study addresses this gap by proposing and testing a "Two-Pathway Heuristic Model" of impression formation, arguing that users evaluate profiles via two parallel pathways: an immediate, affective assessment of facial attractiveness and a more inferential assessment of social attractiveness (or vibe). Using data from 10,619 users on a major heterosexual dating platform in South Korea, we leverage a Large Multimodal Model (LMM) to quantify facial attractiveness and social attractiveness, decomposing the latter into its key components: social, economic, and cultural capital. Through econometric analyses, we examine both the independent and the interaction effects of these two pathways on matching success. Our findings reveal that while both facial attractiveness and social attractiveness are strong predictors of matching success, the effect of facial attractiveness is particularly decisive for male profiles. Furthermore, our analysis uncovers significant negative interaction effects, suggesting that when a user's facial attractiveness is already very high, adding strong signals of social or cultural capital yields diminishing returns. These results offer novel insights into the non-linear patterns of impression formation in digital environments and provide practical implications for both users and platform designers, suggesting a need for interfaces that showcase multi-dimensional aspects of identity beyond physical appearance.

Online matching platforms constrain users to static profiles, producing a mismatch between the idealized self a user presents and the heterogeneous preferences of potential partners. Drawing on self-discrepancy theory, we conceptualize this mismatch as an interpersonal gap between one's presented self and what each partner desires to see, with AI serving as a mediator to resolve it. Following the design science perspective, we propose VISAGE, an AI system comprising two artifacts grounded in distinct human-AI collaboration principles. The augmentation artifact selects optimal images from users' existing assets, whereas the assemblage artifact generates new images tailored to individual partner preferences. Using data from a major online dating platform, we evaluate VISAGE at both the user and platform levels. At the user level, model-predicted ratings suggest that both artifacts improve attractiveness ratings. Their relative effectiveness varies with partner-preference heterogeneity and user impression management skill. At the platform level, agent-based simulations suggest that VISAGE enhances matching efficiency and reduces inequality in matching opportunities, although optimal deployment strategies depend on the platform's recommendation algorithm.

WITS 2024, Bangkok · ICIS 2022, Copenhagen · WITS 2021, Austin.

Online learning platforms have begun to replace human instructors with AI-generated instructors that present complete courses on screen. Whether this substitution supports actual learning behavior remains unsettled, because instruction bundles two different resources: the transmission of knowledge and the provision of emotional support. Drawing on mind perception theory, we propose that AI instruction fails selectively where learning consumes a feeling mind, and that it reaches learners through two separable representation layers: an identity cue that changes who learners believe the instructor is, and an embodied artifact that changes what learners see and hear. We examine this account with behavioral records from a large online learning platform (11,800 user–lecture observations from 2,784 matched users) and a three-condition randomized experiment (N = 298). In the field, the average AI–human difference in lecture completion was indistinguishable from zero, but the difference was 10.9 percentage points more negative in affective than in cognitive courses. In the experiment, labeling an otherwise identical human-delivered video as AI left perceived naturalness intact but reduced relational appraisals of the instructor and willingness to pay (−29%), whereas replacing the video with an embodied AI artifact lowered every appraisal.

CIST 2024, Seattle.

Smaller independent hotels lack the occupancy data that large chains use for dynamic pricing. The paper proposes CAMP (Competitor-based Accommodation Market Pricing), which identifies competitors through hierarchical clustering and predicts rates from competitor prices and external signals such as regional search trends, weather, and economic indicators.

With the increasing use of online matching platforms, predicting matching probability between users is crucial for efficient market design. Although previous studies have constructed various visual features to predict matching probability, facial features, which are important in online matching, have not been widely used. We find that deep learning-enabled facial features can significantly enhance the prediction accuracy of a user's partner preferences from the individual rating prediction analysis in an online dating market. We also build prediction models for each gender and use prior theories to explain different contributing factors of the models. Furthermore, we propose a novel method to visually interpret facial features using the generative adversarial network (GAN). Our work contributes the literature by providing a framework to develop and interpret facial features to investigate underlying mechanisms in online matching markets.

₩30,000,000

Principal investigator

I analyzed the effects of generative AI adoption on digital platforms and derived utilization strategies from the results.

₩140,000,000

Lead research associate

I developed the deep learning model that classifies 3D assets for the accumulation and sharing platform.

₩56,800,000

Lead research associate

I developed the deep learning demand forecasting model and prepared it to run as a SaaS service rather than a one-off analysis.

₩30,000,000

Lead research associate

I developed the assessment tool that measures the ministry's digital innovation capability.

₩102,000,000

Research associate

I analyzed the advertising data and developed the matching engine that pairs advertisers with influencers.