Abstract
E-Commerce personalization aims to provide individualized offers, product recommendations, and other content to customers based on their interests. The foundation of any personalization effort is customer segmentation. The idea of customer segmentation is to group customers together according to identifiable segmentation attributes including geolocation, gender, age, and interests. Personality quiz turns out to be a powerful tool that enables costumer segmentation by actively asking them questions, and marketers are using it as an effective method of generating leads and increasing e-commerce sales. In this paper, we study the problem of how to select and sequence a group of quiz questions so as to optimize the quality of customer segmentation. In particular, we use conditional entropy to measure the utility of a given group of quiz questions. Our objective is to compute a sequence of quiz questions that lead to the maximum utility. We model the user behavior when interacting with a sequence of quiz questions as a Markov process. Then we develop a series of question selection and sequencing strategies with provable performance bound.
Introduction
In modern e-commerce, understanding the customer is paramount. Personality quizzes have emerged as a popular mechanism for "profiling" users, allowing platforms to deliver tailored recommendations. However, traditional research often assumes a "captive audience" model, where users are expected to complete all presented questions. In practice, user engagement is fragile. This paper addresses the reality of the Cascade Browse Model, where users interact with questions sequentially and may choose to answer, skip (via a "Prefer Not to Answer" or PNA option), or abandon the quiz entirely at any stage.
The challenge lies in the joint selection and sequencing of questions. Designers must select the most informative questions while ordering them to minimize attrition and maximize cumulative utility. Given that information utility typically exhibits diminishing returns—a property known as submodularity—the authors frame this as a Cascade Submodular Maximization problem.
Mathematical Framework
The authors model the user's journey as a Markov process. Each question $q$ is characterized by its answer-through-rate ($p_q^+$) and its continuation probability ($c_q$), which represents the likelihood of a user moving to the next question after the current interaction. The goal is to maximize the expected utility $f(Q)$ of a sequence $Q$.
The notation used to define the reachability of questions and the aggregated continuation probabilities is summarized in Table 1. This formalization allows the problem to be treated as an optimization over sequences where position bias significantly impacts the expected utility.
Methodology
To address the NP-hard nature of joint selection and sequencing, the authors propose several non-adaptive algorithms. These strategies do not change the quiz sequence in real-time but are optimized based on prior knowledge of question performance metrics.
Algorithmic Innovations
- Reachability Thresholding: The researchers demonstrate that questions with very low "reachability" (the probability of being seen by the user) can be pruned from the candidate set with minimal impact on total utility.
- Problem Transformation: The complex sequencing task is transformed into a submodular maximization problem subject to linear constraints. By iterating through potential "terminal" questions and optimizing the preceding set, the authors derive a provable approximation ratio.
- Virtual Question Mapping: For more complex versions (like slot-dependent decay), the framework utilizes "virtual copies" of questions for every possible position, effectively converting the sequencing problem into a selection problem over an expanded set.
Experimental Evaluation
The proposed methods were evaluated using extensive numerical simulations.
- Setup: 1,000 synthetic problem instances were generated, featuring quizzes with 12 to 200 candidate questions.
- Metrics: The primary utility metric was the expected entropy reduction of customer attributes.
- Performance: The proposed Question Selection and Sequencing (QSS) algorithm significantly outperformed standard heuristics. While the theoretical worst-case approximation ratio is approximately 0.158, empirical results showed QSS consistently achieving over 87% of the true optimum. It outperformed random sequencing by 10x and exceeded "MaxEnt" strategies (which ignore abandonment probabilities) by 10-20%.
Discussion and Insights
A key finding of the study is the PNA Paradox: while offering a "Prefer Not to Answer" option might seem to reduce data collection, its absence often leads to higher abandonment rates, ultimately yielding less information. Furthermore, the results suggest that order often matters more than content. High-utility but "sensitive" questions should be placed later in the sequence to build user momentum with "safer" questions first.
Limitations
The current work focuses on non-adaptive sequences. While adaptive quizzes—where subsequent questions depend on previous answers—could offer higher utility, they present significantly greater analytical and implementation complexity. Additionally, the model assumes known parameters for $p^+$ and $c$, which in real-world scenarios must be estimated from historical data.
Conclusion
This paper provides the first algorithmic framework with provable performance guarantees for the Cascade Submodular Maximization problem. By bridging the gap between assortment optimization and information theory, the authors offer a robust toolkit for designing online quizzes that maximize customer insights in the face of user attrition.
Summary updated