Ph.D. in Economics, University of Pennsylvania, 2017 - Now
B.S. in Mathematical Decision Science & Economics, University of North Carolina at Chapel Hill, 2014 - 2017
Research Interests (influencer marketing, social media, digital marketing)
My research focuses on developing both theoretical and quantitative models to study unique aspects of influencer marketing, given the multi-level heterogeneity in influencers, content, and browsing preferences on social media. They provide managerial insights on how firms could 1) optimize their influencer selection, and 2) design more effective influencer marketing content. In particular, beyond modeling the strategic interactions between firms, influencers, and users, I focus on applying and proposing new machine learning methods, such as deep generative models, causal ML, and Bayesian econometrics, to empirically test and investigate interesting marketing questions stemming from that heterogeneity, driven by consumer and diffusion theories.
Macro or Micro? Optimal Influencer Selection by Follower Elasticity (Zijun Tian, Ryan Dew and Raghuram Iyengar)
Despite the increasing practices and research in influencer marketing, there is still an ongoing debate between macro vs. micro influencers in terms of their profitability. Macro influencers can leverage their social influence to generate larger content reach for firms, while micro influencers have closer connections with their followers and are less costly to sponsor given the standard follower-based payment in the industry. To solve this trade-off, we need to quantify the true returns to followers in terms of reach. So, in this paper, we estimate the causal impact of follower size on video impression based on massive video data from TikTok. However, it is challenging for three reasons: 1) multiple endogeneity (from both observables and unobservables); 2) the unstructured and multi-modal video content to be learned and included as meaningful and efficient controls; 3) heterogeneous and nonlinear treatment effects. To address them, we adopt a Deep IV approach to estimate the causal relationship between follower size and video impression, with the latent content controls obtained from our proposed multi-modal feature extraction algorithm and representation learning framework (VS-VAE). We show that the treatment effect (i.e. follower elasticity), on average, exhibits an "inverted U-shape'', leading to an S-shaped impression growth trajectory by followers. However, both its shape and magnitude vary significantly by content. We then not only derive interpretable insights on which influencers are more effective in driving impression for which content but also make content-based influencer recommendation to firms based on the heterogeneous treatment effects. Finally, we provide counterfactual studies to illustrate the dramatic increase in profits under our optimal influencer choice and note some novel observations that differ influencer marketing from other advertising.
How Should Firms Cooperate with Honest Influencers? (Zijun Tian)
In this paper, I strategically model the firm’s cooperation with an honest influencer who is truthful in his/her review. In particular, I assume that the influencer’s review is normally distributed around the true product quality yet endures two sources of uncertainties: 1) imperfect learning of the product; 2) variance in the writing. However, the firm can improve the informativeness of the influencer’s review by revealing the true product quality to the influencer, thus eliminating his/her learning noises. To model the influencer's concern of his/her reputation, I set his/her utility function to be a linear combination of the firm's payment and the negative amount quality inflation perceived by the "misled'' consumers. I find that the firm can still achieve a positive profit even when the product is of fairly lower-than-expected quality since it can asymmetrically gain from the uncertainties introduced through the influencer’s message. Moreover, analogous to the “signal jamming” literature in persuasion, I find that with very favorable or unfavorable product quality, the firm does not distort the influencer’s signal but reveals the true quality to the influencer. However, it should lower the influencer’s review precision by not revealing the true quality when its product is mediocre. Finally, I consider the heterogeneity in firms by classifying them into new vs. established who differ in the variance of their product quality. New firms with larger quality variance benefit more from noisier reviews over a larger range of product qualities and generally achieve higher profits from influencer sponsorships. The findings can be generalized to new vs. mature products as well, suggesting ways that firms can profit more from their cooperation with honest influencers, e.g. sponsoring products from new product lines or cross-brand collaboration.
Research in Progress
What Makes a Hashtag Engaging: Analyzing Topic Design and Evolution on TikTok (Zijun Tian, Ryan Dew and Raghuram Iyengar)
In this project, we investigate the unique form of influencer marketing via sponsored topics on TikTok (commonly found on Twitter's and Weibo's Trend page as well), which we call the "campaign hashtags", and explore sources of variations that drive their success on engaging users to browse (i.e. popularity) and participate (i.e. content variety). Specifically, a campaign hashtag is consisted of a name, a short description of what it is about, and a set of videos shown on its main page created first by the sponsored influencers (i.e. the seeding videos) followed by the public. We hypothesize that all three could be associated with the total amount of attention and organic content creation under a hashtag and develop machine learning algorithms to automatically 1) get embedding-based unexpectedness measure for each hashtag, 2) classify hashtags into substantive topics, 3) quantify the amount of imitation/innovation across the top videos, and 4) predict seeding video popularity and examine their ranking optimality. We find that the unexpectedness of a hashtag exerts a significant U shaped/inverted U-shaped effect on its popularity/content variety. Higher early content-creating conformity helps a hashtag become more popular, which is more prominent under socializing and narrower-scoped hashtags. For campaign hashtags specifically, ranking seeding videos based on our predicted popularity has shown to significantly improve their aggregate popularity.
Leveraging Free Advertising from the Amateur Influencers on Instagram (Zijun Tian)
On social media, influencers might differ in their status by which we classify them into the “professionals” who are already established in their niches to get sponsorships from firms, and the “amateurs” who are not there yet. In order to profit from their posts sooner, the amateurs might choose to imitate what the professionals post. Specifically, if a professional influencer posts a sponsored review for a product, then the amateurs might follow by reviewing the same product themselves, with the hope that they might get their similar reviews exposed to the professional’s large audience under the targeting algorithms on social media. For firms, these imitating reviews from the amateurs act as free advertising in addition to their original paid reviews, seeding additional eMOM among the amateurs' followers. However, to leverage such free advertising by encouraging more imitation from the amateurs, firms need to understand what they tend to imitate: the popular topics, the successful idols, or the surprising products? In this project, I strategically model the incentives behind the amateur’s imitation and empirically test whether the amateurs indeed imitate, when they benefit from such imitation, what they imitate, whether they alternate between or stick to one imitating strategy, and how the above would differ by influencer niches, based on panel posting dataset for a large, random set of professional and amateur influencers on Instagram. In particular, I design imitating scores that quantify the intensity of each imitating strategy used in each influencer’s post by comparing post similarity from different dimensions via advanced machine learning methods. The results would shed light on whether firms should sponsor more influencers, (if the amateurs imitate popular topics), more influential influencers (if they imitate successful idols), or influencers from different niches (if they imitate surprising products), and how these decisions differ by niche categories.
Heterogeneous Preferences toward Beauty Videos: An Interpretable Deep Learning Approach with Multi-level Attentions (Zijun Tian)
Firms want influencers to create content built around the preferences of their target audience. Analogous to product attributes that segment the market demand of a product across consumers, there should be a set of underlying content attributes that relate to and differentiate the browsing preferences across users on social media. Current tools (e.g. Facebook’s Creator Studio) have already suggested creators to test their video performance over some simple attributes (e.g. captions) within their audience. However, given the high-dimensionality of videos, such testable attributes are endless and can defined at different hierarchies. In addition, those potentially unknown to the researcher may more significantly impact user preferences. Therefore, to find them in a data-driven yet still semantically interpretable way, in this project, I design an interpretable deep learning approach with multi-level attentions that automatically learns the latent user preferences toward a latent set of video attributes. Attention fusion happens both at the demographics level (in the user network) and at the frame and modality level (in the video network) whose weights introduce interpretation to the learned preferences and video attributes. We apply the approach to beauty videos (one of the most popular influencer niches) in particular, with the hope that we discover unobvious preference heterogeneity and cross-modality attributes beyond typical user groups (e.g. gender) and simple video features (e.g. captions). The results can be easily generalized to other industries, helping firms create more preferable (or test more effective) content toward their target audience on social media.