Research Interest

My research bridges AI, economics, and consumer behavior, focusing on algorithmic scoring of visual signals, peer-to-peer marketplaces, and sustainability certification on digital platforms. Trained in Carnegie Mellon’s quantitative marketing and behavioral science tradition, I combine empirical modeling with interdisciplinary methods. My long-term interest in sustainability is rooted in early exposure to large-scale water resource conservation and allocation projects through my family’s work, which shaped my perspective on aligning digital innovation with real-world environmental systems. This unique background inspires my vision to build global collaborations connecting research, industry, and sustainability initiatives.

Job Market Paper

  • Xiaohang (Flora) Feng.

    Working paper available. [Link]

    Abstract: Sustainability badges are increasingly central to e-commerce, yet their long-term effects on platform outcomes remain unexplored because A/B tests are typically short-term. Although badges attract green consumers, price premiums and dissatisfaction gaps can ultimately reduce overall demand. This study evaluates Amazon’s Climate Pledge Friendly (CPF) badge using a dynamic structural model integrated with a multimodal generative AI framework. We ask: (1) What badge threshold best supports long-term platform goals? (2) Should the badging strategy change as the proportion of green products varies? (3) How can generative AI address the challenges of unstructured data in policy simulations? Using data on more than 1,200 products tracked for six months in Amazon.com, the structural model captures forward-looking seller behavior, where daily certification and pricing decisions account for future returns. To simulate multimodal content adaptation, we introduce the Green-Flora-GPT framework, which transforms non-green product listings into green-aligned images and descriptions. The model uses a novel utility-aware contrastive loss to maintain semantic coherence, brand consistency, and authentic green signaling. We find that full badge coverage is not optimal: the median revenue of badged products peaks when 80% of products are certified, whereas consumer welfare is highest at a much lower badge rate (20%). As more products become green, the optimal strategy for platform-wide total revenue becomes increasingly selective, although overall badge visibility remains relatively stable (10–30%). This study offers practical insights for managing sustainability signaling in evolving digital marketplaces.

    Keywords: sustainability, eco-labeling, e-commerce, seller competition, market equilibrium, structural model, generative AI, multimodal representation learning

Accepted Papers

  • Xiaohang (Flora) Feng, Shunyuan Zhang, Xiao Liu, Kannan Srinivasan, and Cait Lamberton.

    Accepted at Journal of Marketing Research. 2025. [Link]

    Abstract: It has long been a mantra of marketing practice that, particularly in low-involvement situations, spokespeople should be physically attractive. This article suggests there is a higher probability of gaining fame and influence (i.e., celebrity potential) than is captured by attractiveness or typicality. The authors identify 11 facial features that may predict celebrity potential by virtue of their purported relationship with charisma and resulting personality trait inferences. Using machine learning methods and a sample of 22,000 faces, the authors calculate the direction and strength of the correlation of each feature with celebrity potential. The model is 95.92% accurate in predicting whether a given face belongs to a celebrity or noncelebrity, and it enables a celebrity visual potential (CVP) metric to be calculated for any face. Two controlled experiments and two studies using photographs of faces of Instagram and LinkedIn users further validate that the model-generated CVP is consistent with human-rated CVP, showing predictive power above and beyond facial typicality and averageness. This research challenges prior assumptions about the importance of attractiveness in spokesperson choice, offers a useful additional metric for marketers, and provides novel insights about the relative importance of various inferred personality traits for celebrity potential.

    Keywords: celebrity visual potential, facial features, personality traits, deep learning, explainable artificial intelligence

  • Xiaohang (Flora) Feng, Charis Li, and Shunyuan Zhang.

    Accepted at Journal of Consumer Research. 2025. [Link]

    Abstract: Peer-to-peer (P2P) marketplaces have seen exponential growth in recent years, featuring unique offerings from individual providers. However, scalable quantification of visual uniqueness and their impacts on platforms like Airbnb remain largely unexplored. We address this gap by developing, validating, and applying an unsupervised machine learning model to automatically extract uniqueness from images and quantify its impact on demand. We first construct a machine learning model, informed by cognitive psychology, to assess visual uniqueness in 481,747 property images, achieving high accuracy and interpretability. Next, we validate our model through three studies involving various participant populations and methods, confirming that the model’s predictions of visual uniqueness align with human judgment. Finally, we apply this model to demand data of Airbnb properties in New York City spanning 13months. We find an inverted U-shaped relationship between visual uniqueness and demand, with two significant moderation effects: properties with higher response rates or overall ratings benefit more from visual uniqueness. This research provides valuable insights for P2P platforms like Airbnb, highlighting the strategic use of visual uniqueness to enhance visual appeal and market performance. It also offers a new methodological roadmap for integrating psychological insights into the development and validation of unsupervised machine learning models.

    Keywords: visual uniqueness, Airbnb, unsupervised contrastive learning, interpretable machine learning, image analytics, peer-to-peer marketplace

Working Papers

  • Xiaohang (Flora) Feng, Xiao Liu, Shunyuan Zhang, and Kannan Srinivasan.

    Revise & Resubmit at Marketing Science. [Link]

    Abstract: Amazon introduced the Climate Pledge Friendly (CPF) badge by consolidating various green certificates to examine its impact on market dynamics. We applied a game-theoretic model and causal inference using data from Amazon.com to explore the effects of this badge on consumer behavior, seller pricing, and market concentration. Our theoretical model outlines a three-stage process where sellers set prices, the marketplace determines badge eligibility, and consumers make purchase decisions. We discovered that increased demand, higher prices, and reduced market concentration occur when the benefits gained from attracting green consumers exceed the detriments from alienating non-green consumers due to increased prices. Optimal conditions were identified where certifying only the most sustainable products maximizes outcomes over strategies that result in either all or no products being badged. Empirically, we gathered six months of data on 6,606 products across eight categories, using the interactive fixed effect counterfactual (IFEct) estimator to manage endogeneity and treatment reversals. Our findings indicate that the CPF badge significantly enhances sales volume, increases product prices, and decreases market concentration. These results guide sellers considering green certification and platforms contemplating unified green badge policies.

    Keywords: sustainability, eco-labeling, e-commerce, seller competition, causal inference, multimodal vector representation

  • Xiaohang (Flora) Feng, Yiling Xie, and Jehoshua Eliashberg.

    Working paper available.

    Abstract: Movie trailers are information‑intensive, combining audio and visual cues that strongly influence a film’s commercial success. Yet little research addresses how to automatically design trailers that maximize economic value at scale. Using a dataset of 352 trailers, we apply advanced analytics in three stages. First, we extract editable visual and audio signals from unstructured video with state‑of‑the‑art machine‑learning models. We then merge two visual signals (brightness and RMS contrast) and two audio signals (loudness and pitch) into latent factors via dynamic factor analysis. Second, we evaluate how these dynamic factors predict trailer ratings with functional regression. Results show that audio has a consistently positive effect during the first half of a trailer but a mixed effect later, whereas visual signals exert a positive influence for most of the runtime. Finally, we develop a scalable framework that designs an optimal audio trajectory, conditional on the visual path, using optimal–control theory. We compute the mean absolute percentage error (MAPE) between each trailer’s original and optimal audio track and regress this gap on trailer characteristics. Only rating and the adventure/drama genres significantly explain MAPE, indicating that higher‑rated trailers already align closely with the optimal audio pattern. The framework offers actionable guidance: producers can diagnose and refine audio–visual balance to boost viewer utility. Because the pipeline relies on general video features and control logic, it extends beyond trailers to broader video‑production and generative‑AI applications.

    Keywords: movie trailers, audio–visual signals, dynamic factor analysis, functional regression, optimal control, machine learning, video analytics

Book Chapters

  • Xiaohang (Flora) Feng, Shunyuan Zhang, and Kannan Srinivasan.

    Artificial Intelligence in Marketing. Edited by Naresh K. Malhotra, K. Sudhir, and Olivier Toubia. 2023. [Link]

    Review of Marketing Research. Emerald Publishing Limited.

  • Xiaohang (Flora) Feng, Shunyuan Zhang.

    Elgar Encyclopedia of Pricing. Edward Elgar Publishing. 2024. [Link]