Market segmentation in online platforms

Franco Berbeglia, Gerardo Berbeglia, Pascal Van Hentenryck

Abstract

In digital trial-offer markets, platforms utilize ranking algorithms and social signals to guide consumer choice. While traditional models suggest that highlighting popular items improves market efficiency, this research demonstrates that such strategies can be counterproductive when consumer preferences are heterogeneous. By employing Mixed Multinomial Logit (MMNL) modeling and agent-based simulations, this paper explores how global popularity signals can lead to a "popularity trap," reducing total sales. The authors propose market segmentation as a solution, proving that tailoring rankings to specific consumer groups can significantly enhance platform performance.

Introduction

Online platforms such as Netflix, the App Store, and Steam operate as "trial-offer" markets. In these environments, consumers typically "try" a product—by watching a trailer, reading a description, or playing a demo—before deciding whether to "buy" or commit to it. Platforms influence these decisions through two primary levers: position bias (placing items at the top of a list) and social influence (displaying popularity metrics like download counts or view totals).

Historically, research into these dynamics assumed a "homogeneous" consumer base, where users share similar tastes. Under such assumptions, the optimal strategy is straightforward: rank products by quality and display their popularity to accelerate discovery. However, this paper addresses the more complex reality of heterogeneous preferences. When tastes diverge significantly, a global popularity signal may inadvertently steer users toward products they will not enjoy, leading to wasted trials and lower overall conversion rates.

Methodology

The researchers utilized a multi-faceted approach to analyze market dynamics under preference heterogeneity:

  • The Model: The study employs the Mixed Multinomial Logit (MMNL) model. This framework is highly flexible, capable of approximating any random utility model by incorporating intrinsic product appeal, the influence of ranking positions, and the impact of cumulative past purchases.
  • Theoretical Analysis: The authors examined the asymptotic behavior of the market to determine if it converges toward a "monopoly" state where a single product dominates the market share.
  • Computational Complexity: The study investigated the algorithmic difficulty of optimizing rankings in real-time for diverse user bases.
  • Simulation: To validate theoretical findings, the authors conducted agent-based simulations emulating the famous MusicLab experiment. The simulation involved 50 items and 200,000 consumer arrivals across four distinct scenarios where the relationship between a product's initial appeal and its actual quality varied.

The Challenge of Heterogeneity

The "Popularity Trap"

The core finding of the paper is that global popularity signals can be "detrimental" in diverse markets. If a product is highly popular within one segment but low-quality for another, a global signal may persuade the second group to try it. Because the product does not align with their tastes, they are unlikely to purchase it. This creates a "popularity trap" where the platform's limited trial opportunities are squandered on poor matches, reducing the total expected purchase rate.

Computational Barriers

The authors prove that finding the "Performance Ranking"—the specific ordering of products that maximizes the probability of the next purchase—is NP-hard when there are two or more consumer segments. This suggests that for real-world platforms, finding a mathematically perfect ranking in real-time is computationally infeasible, necessitating the use of smart heuristics and segmentation strategies.

The Power of Market Segmentation

To mitigate the negative effects of the popularity trap, the authors propose the Segmented Quality and Segmented Social Influence (SQSSI) policy. This approach involves showing users a ranking and a popularity signal specific to their identified segment (e.g., "Top 10 for Strategy Gamers").

Key results from the analysis of segmentation include:

  • Efficiency Gains: In simulations, the SQSSI policy consistently outperformed global ranking strategies. In certain scenarios, segmentation was nearly twice as efficient as using a global popularity signal.
  • Scaling with Segments: Theoretically, the platform can increase expected purchases by a factor of $K$, where $K$ is the number of distinct consumer segments.
  • The Risk of Global Signals: If consumer preferences are sufficiently different, showing a global popularity signal can actually make a platform perform $K$ times worse than showing no social signal at all.

Market Convergence

Despite the complexities of diverse tastes, the study found that markets still tend toward predictable long-term outcomes. If a platform uses an "Average Quality" ranking, the market will eventually be dominated by the product that possesses the highest weighted average quality across all segments. This suggests that while short-term dynamics are messy, the long-term market structure remains governed by underlying quality metrics.

Conclusion and Future Directions

This research highlights a critical tension in platform design: the trade-off between the simplicity of global signals and the efficiency of personalized rankings. While market segmentation offers a clear path to higher sales, its success depends on the platform's ability to accurately identify a user's segment. The authors note that their findings remain robust even with minor misclassification errors (up to 10-20%), but real-world implementation requires sophisticated user profiling.

Future research may explore non-linear "herding" effects, where social influence does not scale linearly with past purchases, and the dynamics of evolving or overlapping consumer segments. Additionally, extending the model to consider consumer welfare and the potential for strategic manipulation of reviews remains a vital area for further study.

Summary updated

BibTeX
@article{berbeglia2021market,
  author = {Franco Berbeglia and Gerardo Berbeglia and Pascal Van Hentenryck},
  title = {{Market segmentation in online platforms}},
  journal = {Eur. J. Oper. Res.},
  year = {2021},
  doi = {10.1016/j.ejor.2021.03.056},
}
Figure