Towards a theory model for product search

Beibei Li, Anindya Ghose, Panagiotis G. Ipeirotis2011 · DOI

TL;DR

This paper proposes a novel product search ranking model grounded in economic utility theory, moving beyond traditional information retrieval. By estimating consumer surplus, it generates rankings that better reflect "value for money," significantly outperforming existing methods in user studies.

What Problem Does This Paper Address?

Online product search engines currently rely heavily on information retrieval models or simple ranking criteria like price or popularity. However, the decision-making process for purchasing a product is fundamentally different from finding a document. Consumers don't just seek relevance; they aim to identify the "best deal" that aligns with their specific, often multidimensional, preferences. Existing systems struggle with this because they often ignore consumer heterogeneity, fail to leverage rich online community data beyond simple ratings, and require users to log in or provide personal information, which many are reluctant to do. This leads to suboptimal search results that don't truly capture what users are optimizing for.

Methods

The core of this paper's contribution lies in its theoretical framework and estimation methodology.

  • Economic Foundation: The model is built on expected utility theory from economics. It posits that consumers aim to maximize their utility, which is a function of both the product's characteristics and the utility lost from spending money.
  • Utility and Surplus:
    • Product Utility ($U_p(X)$): Modeled using Lancaster's characteristics theory and Rosen's hedonic price model, where a product's utility is a weighted sum of its characteristics.
  • Money Utility ($U_m(I)$): Assumed to be concave and increasing, reflecting diminishing marginal utility of money. For small price changes, a quasi-linear approximation is used.
  • Consumer Surplus (CS): Defined as the difference between the utility gained from a product and the utility lost from its price. This surplus is the key metric for ranking.
  • Demand Estimation: To overcome the challenge of unobservable individual preferences, the authors employ demand estimation techniques from economics.
  • Homogeneous Consumers (Logit Model): Initially, they model a scenario where all consumers have the same preferences. Using aggregate demand data (market shares), they estimate the weights ($\beta$) for product characteristics and the sensitivity to price ($\alpha$) via logistic regression.
  • Heterogeneous Consumers (BLP Model): Recognizing that preferences vary, they extend the model using the Berry-Levinsohn-Pakes (BLP) model. Here, consumer preferences ($\beta$ and $\alpha$) are modeled as functions of demographic variables (e.g., age, travel purpose) and income. This allows for personalized rankings. The estimation involves an iterative process using instrumental variables and generalized method of moments (GMM) to find parameters that best explain observed market shares.
  • Ranking: Products are ranked by their estimated consumer surplus. For personalized rankings, the surplus is calculated based on a specific consumer's demographic profile.

Datasets

  • Travelocity Transactions: A dataset of hotel booking transactions for 2117 hotels in the US from November 2008 to January 2009 was used to compute market shares.
  • TripAdvisor Data: Used to extract hotel service characteristics (value, room, location, cleanliness, service, check-in, business service) and demographic information of reviewers (travel purpose, age group).
  • Geo-mapping and Social Geotags: Used to identify external amenities and public transportation near hotels.
  • Image Classification & Mechanical Turk: Used to identify features like nearby beaches or downtown areas.
  • Online Reviews: Text from reviews was analyzed for stylistic characteristics (subjectivity, readability) and reviewer identity disclosure.

Key Insights and Ideas

  • Economic Theory as a Foundation for Search: The paper demonstrates that microeconomic principles like utility maximization and consumer surplus can provide a powerful, theory-driven framework for product search, moving beyond purely algorithmic approaches.
  • "Value for Money" is Key: The core insight is that users optimize for "value for money," which is precisely what consumer surplus aims to capture. This is more nuanced than just price or rating.
  • Privacy-Preserving Preference Inference: A significant contribution is the ability to infer consumer preferences from aggregate demand data, without needing individual-level purchase histories or explicit demographic information from users. This addresses privacy concerns.
  • Heterogeneity Matters: The BLP model's success highlights that accounting for consumer heterogeneity (demographics, income) is crucial for generating truly effective personalized recommendations.
  • Causal Model for Dynamic Ranking: The model is "causal," meaning it can predict how changes (e.g., new products, price changes, feature updates) will affect rankings, enabling dynamic adjustments without waiting for new demand data.

Key Results

  • Overwhelming Preference for Surplus-Based Rankings: Extensive user studies involving over 15,000 comparisons showed an "overwhelming preference" for the rankings generated by their surplus-based techniques compared to numerous state-of-the-art baselines.
  • Personalized Rankings Outperform General Rankings: Personalized surplus-based rankings, tailored to specific demographic profiles (e.g., business traveler), were significantly preferred over general surplus-based rankings.
  • Diversity as a Natural Outcome: The surplus-based approach naturally leads to diverse results, offering a mix of price and quality, which users appreciated over artificially introduced diversity metrics.
  • Feature Importance: Analysis revealed that "beach" had the highest positive impact among location characteristics, while factual reviews were preferred over subjective ones. "Review value" ratings surprisingly had a negative sign, suggesting these hotels might be "undiscovered."

Limitations and Future Work

  • Rationality Assumption: The model assumes consumers are rational utility maximizers, which is a simplification of real-world behavior. The authors acknowledge this but note that even imperfect theories can yield good results.
  • Data Sparsity: Some demographic fields were sparsely populated, limiting their direct use in the model.
  • Future Work:
    • Incorporate browsing history and learning behavior to model "satisficing" versus "utility optimizing" search strategies.
  • Investigate the conversion rate differences when users are presented with surplus-based rankings.
  • Extend the model to handle product bundles.
  • Further explore the "causal" nature of the model to predict market dynamics.

Summary updated

Abstract

With the growing pervasiveness of the Internet, online search for products and services is constantly increasing. Most product search engines are based on adaptations of theoretical models devised for information retrieval. However, the decision mechanism that underlies the process of buying a product is different than the process of locating relevant documents or objects.
BibTeX
@misc{li2011towards,
  author = {Beibei Li and Anindya Ghose and Panagiotis G. Ipeirotis},
  title = {{Towards a theory model for product search}},
  year = {2011},
  doi = {10.1145/1963405.1963453},
}
Figure