Han Chen

Han Chen

Doctoral Candidate, Marketing

CV | GitHub | LinkedIn


I am a Ph.D. candidate in marketing. I employ econometrics and causal inference, lab/field experiments, and machine/deep learning to solve marketing problems in the domains of digital advertising, price promotion, product innovation, and human-algorithm interaction. Broadly, my research examines 1) how firms effectively use marketing mix tools to influence consumer behavior in the era of artificial intelligence and the digital world, and 2) how marketers use machine learning algorithms to boost the efficacy of marketing mix tools. My dissertation and research revolve around (1) the substantive domain of consumers’ behavioral responses to marketing messages and disclosures, with a special focus on their sequential, dynamic, and unintended effects, and (2) the optimization of firms' sequential marketing communications and interventions. I am on the 2022/2023 academic job market.


PhD Candidate, Marketing (Expected 2023)

    Temple University, Fox School of Business, PA, USA
    Dissertation Committee: Yang Wang (co-chair), Joydeep Srivastava (co-chair), and Subodha Kumar

Master, Management Science and Engineering (ranked 1/31 in class), 2017

    Northeastern University, China
    Thesis: Measurement and Evaluation of User Interactive Experience and Product Design Implications: A Multi-Method Approach Applied to Mobile Shopping Apps

Bachelor, Industrial Engineering (ranked 3/62 in class), 2014

    Northeastern University, China
    Thesis: Optimal Logistics Center Location: Iterative Center-Of-Gravity Method and Application


Dissertation and Research

Flying High, Landing Low? Effect of Tensile Price Promotions on Store Traffic and Sales [Paper] [Simulation Code]

This research examines the effectiveness of three types of tensile discounts in advertising storewide promotions: a maximum discount (i.e., up to Y% off), a range of discounts (i.e., X% off to Y% off), and a minimum discount (i.e., starting at X% off) on consumers’ initial perceptions and expectations, store visits, and purchases. Using multiple methods, including lab studies, a quasi-field study, empirical analysis of data from a large-scale storewide promotion campaign, and numerical simulations, the findings suggest that while a maximum tensile discount is most effective in shaping consumers’ initial perceptions, and thus store visits, it does not necessarily boost purchases in some conditions. Our findings are consistent with the reasoning that because a maximum tensile discount leads to the most favorable discount expectations, the likelihood and the extent of experiencing negative expectation disconfirmation (i.e., realized discount lower than the expected discount) is highest with a maximum tensile discount, thus discouraging purchases and anticipated satisfaction. The lab studies are supplemented with empirical data and numerical simulations to highlight the double-edged nature of using a maximum discount to advertise storewide promotions. There is thus the need to balance raising expectations of discount and the likelihood of consumers experiencing negative expectation disconfirmation.

This is the most recent version of this paper. It is the first project and dissertation essay I began to work on and among those to which I have devoted the most time (so far). Compared to previous marketing promotion literature, research on storewide sales is challenging as storewide price promotion is commonly used when the entire assortment of products across different categories and price ranges is offered on sale. Thus, above and beyond traditional lab studies, we need new methods and new data. Acquiring research data (both primary and secondary) can be difficult and the path to it may be unexpected. We are finally able to supplement the lab studies with a quasi-field experiment via created retail websites (another story of battling Prolific), empirical data, and numerical simulations. Any comments or questions are welcome and please address them to hanchen@temple.edu.

Han Chen, Joydeep Srivastava

  • Working Paper

First Week Human, Second Week Algorithm: Curation Source Advertising for Breadth-Then-Depth Selling [Paper]

Many platforms like Netflix strategically advertise and highlight their algorithm curation as a core product benefit for customers, while others like HBO Max and Apple News differentiate on the basis of human curation. Using sequential field experiments on a large mobile reading platform, we examine whether and how advertising algorithm versus human curation impacts consumption behaviors, and how they can be optimally sequenced to best improve the advertising effects. Findings suggest that, despite their similar efficacy in lifting purchases when used in isolation, a human-then-algorithm (HA) advertising sequence surpasses the reversed and non-hybrid sequences’ purchases by 7%–11%. We find advertising algorithm (human) curation leads to depth (breadth) selling by increasing purchases in previously (not) consumed genres. In the best performing HA sequence, this complementarity leads to the initial breadth selling amplifying the subsequent depth selling effectiveness. Importantly, we show that the depth (breadth) selling effects are driven by consumer preference shifts towards previously (not) consumed genres, rather than merely shifting attention to source-labeled curated assortments that happen to be deeper (broader). Our findings suggest that carefully-sequenced advertising of product curation source can help cultivate consumers’ ever-renewing interests by stimulating demand across and within product categories.

This paper was initially conceptualized as a method paper, with the goal of solving for a firm's optimal sequential advertising policy to nudge consumer preference expansion and reinforcement consumption behaviors using deep reinforcement learning (DRL) method. The optimal policy (of when, which ad copy, and to whom) prescribes the advertising trajectory for each individual to maximize their total consumption. Despite the initial plan, the empirical findings turn out interesting and should be the focus of the paper before doing method and optimization. The transition takes time but is necessary in terms of first identifying the causal effect of curation source advertising and telling a clear story about consumer preference expansion and reinforcement. (This is thus another example of a research deviating from what it was initially designed to be.) I recognize the importance of being flexible, and I learn from and enjoy the process of reworking and repositioning this paper from a method-driven to a substance-driven research.

Han Chen, Yang Wang, Hanbing Xue, Yongjun Li, and Xueming Luo

  • Revise and Resubmit at Journal of Marketing Research

Exploring New, Reinforcing Love: Dynamic Product Curations With Quasi-Experiment and Deep Reinforcement Learning

Platforms employ personalized recommendations (e.g., Amazon “Recommended for You”) to expose consumers to a likely preferred subset of the overall product assortment. However, personalized recommendations can thus constrain and concentrate consumer choices to products akin to their previous consumption rather than good outside options, thereby reinforcing consumer prior beliefs and preferences via echo chambers, harming consumer welfare with decreased product variety and quality, and hurting consumer sustained consumption in the long run. To expand consumer preferences by stimulating their adoption of good products from other categories, platforms complement personalized recommendations (PR) with different types of novel recommendations (NR) based on (1) social influence (e.g., Amazon “Popular Items,” Twitter “Trending Hashtags/Topics”), (2) promotions (e.g., Amazon “Deals,” free samples), or (3) product quality (e.g., Amazon “Highly Rated Products”). This paper examines the effect of different types of NR (vs. PR) strategies on consumer consumption, and above and beyond, how should platforms develop a sequential recommendation policy to maximize consumption. Analyzing the reading behavior of over 2.5 million consumers on an e-book platform one month before and after an exogenous technology change that NR and PR panels were highlighted on the first page of the book feed, we find that, despite the minimal immediate effect, the initial NR significantly lift the efficacy of subsequent recommendations on consumer consumption in a longer-term starting from Day 10. However, while social influence- and promotion-based NR have a positive effect relative to PR, surprisingly, the product quality based NR do not. The asymmetric short vs. long run effects of NR suggest platforms have to make intertemporal tradeoffs and be forward-looking to maximize consumer total consumption. We apply a novel deep reinforcement learning (DRL) method to derive an optimal policy (of when, which recommendation, and to whom) that prescribes the recommendation trajectory for each individual to maximize total consumption by cultivating consumer preference expansion and reinforcement behaviors. Compared to classical dynamic programming methods, DRL combines the advantages of reinforcement learning (RL) and deep learning (DL). RL constructs an environment characterized by a Markov decision process (MDP) where an AI agent, without assuming full knowledge of the MDP, learns the policy in trial-and-error by interacting with the counterfactual demand model. It thus converges to the optimal policy via earning while learning. DL handles the massive high-dimensional state-action space in the demand model, easing the curse of dimensionality, and parameterizes and approximates the optimal policy in the agent without explicitly modeling the state transition probability and reward function, reducing the potential model bias and especially advantageous when the exact modeling of MDP is infeasible.

Han Chen, Yang Wang

  • Work in Progress

Algorithmic Ad Creative Design and Dynamic Targeting: A Contextual Multi-Armed Bandits and Deep Learning Approach

Advertisers use new ad creatives to attract consumer attention and alleviate ad wearout effects. However, ad creatives are expensive to produce and their effectiveness for targeted audience need to be tested through reiterated ad experiments before rollout. Luckily, AI is revolutionizing digital advertising by automating new ad creative design and learning the optimal targeting policy on the fly under budget constraint. This paper proposes such an AI ad agent (AIAD) with contextual multi-armed bandits (CMAB) method. It dynamically learns from previous ad performance, creates new ads, and adaptively allocates impressions to different ad creatives in subsequent targeting, thus achieving earning while learning. I test the AIAD using a dataset of over 30,000 ad campaigns run on Alibaba. Campaign ad creatives use different ad message appeals, with diverse contexts of visual design and targeted customer purchase funnel. Descriptive analyses suggest important caveats of different ad appeals across purchase funnel, further necessitating the early learning for maximal earning. The AIAD employs a deep learning demand model, the Linear Thomson Sampling algorithm, and a Doubly Robust estimator. Results of simulations facilitated by an R package suggest that the prescriptive AIAD boosts campaign click rate by 40% compared to the observed level of campaigns without either learning or earning. It also outperforms MAB methods without learning targeting rules by 30%. The AIAD has important implications for digital advertising and targeting.

Han Chen

  • Manuscript in Preparation

Previous Work

From Algorithm Aversion to Appreciation? Dynamic Advertising with Deep Reinforcement Learning [DRL Technical Appendix] [Optimal Policy Visualization]

  • Best Paper in the Big Data and Marketing Analytics Track, 2020 AMA Winter Academic Conference


Department of Marketing 📧: hanchen@temple.edu
Fox School of Busines, Temple University 🌐: hanchenresearch.github.io
Alter Hall A502e, 1801 Liacouras Walk Philadelphia, PA 19122

Last Updated: 07-21-2022

Han Chen