[Paper Review] Market Segmentation Trees

Ali Aouad, Adam N. Elmachtoub+2 more|arXiv (Cornell University)|Jun 4, 2019

Consumer Market Behavior and Pricing10 citations

TL;DR

This paper introduces Market Segmentation Trees (MSTs), a novel interpretable framework that jointly optimizes market segmentation and response modeling by recursively partitioning users based on decision tree splits that maximize predictive accuracy of response models. The method outperforms traditional clustering-response modeling pipelines, achieving superior predictive performance with significantly fewer segments—demonstrated via CMTs for choice modeling and IRTs for bid landscape forecasting—while maintaining computational tractability and interpretability through decision tree structures.

ABSTRACT

We seek to provide an interpretable framework for segmenting users in a population for personalized decision-making. We propose a general methodology, Market Segmentation Trees (MSTs), for learning market segmentations explicitly driven by identifying differences in user response patterns. To demonstrate the versatility of our methodology, we design two new, specialized MST algorithms: (i) Choice Model Trees (CMTs), which can be used to predict a user's choice amongst multiple options and (ii) Isotonic Regression Trees (IRTs), which can be used to solve the bid landscape forecasting problem. We provide a theoretical analysis of the asymptotic running times of our algorithmic methods, which validates their computational tractability on large datasets. We also provide a customizable, open-source code base for training MSTs in Python which employs several strategies for scalability, including parallel processing and warm starts. Finally, we assess the practical performance of MSTs on several synthetic and real world datasets, showing that our method reliably finds market segmentations which accurately model response behavior.

Motivation & Objective

• To address the limitation of traditional market segmentation, which clusters users based on feature similarity rather than response behavior, leading to suboptimal personalization.
• To develop a unified framework that integrates market segmentation and response modeling to improve prediction accuracy in personalized decision-making.
• To ensure interpretability by using decision trees to define market segments, enabling clear, human-readable segmentation rules.
• To demonstrate the method’s scalability and effectiveness on large-scale real-world datasets, particularly in bid landscape forecasting and choice modeling.
• To provide a customizable, open-source implementation with parallel processing and warm starts for practical deployment.

Proposed method

• MSTs use recursive decision tree splits to partition users into market segments based on contextual features, where each split is chosen to maximize the predictive accuracy of the response models fitted in the resulting leaves.
• The training procedure optimizes a global objective that aggregates predictive performance across all segments, ensuring that splits are driven by differences in user response behavior rather than feature similarity alone.
• Two specialized MST variants are proposed: Choice Model Trees (CMTs) for multi-choice response prediction and Isotonic Regression Trees (IRTs) for forecasting win rates in ad auctions using isotonic regression models per segment.
• The algorithm employs a greedy, bottom-up tree-building strategy with pruning to prevent overfitting and ensure generalization.
• The open-source Python implementation supports parallel processing and warm starts to enhance scalability on large datasets.
• Theoretical analysis shows the algorithm’s asymptotic computational complexity is linear in tree depth under mild conditions, supporting tractability on large-scale data.

Experimental results

Research questions

• RQ1 Can a joint market segmentation and response modeling framework outperform traditional two-stage clustering-response modeling approaches in terms of predictive accuracy?
• RQ2 To what extent can decision tree-based segmentation improve interpretability while reducing the number of required market segments?
• RQ3 How does the performance of MSTs compare to state-of-the-art benchmarks in choice modeling and bid landscape forecasting?
• RQ4 Can the proposed method scale effectively to large-scale, real-world datasets while maintaining computational tractability?
• RQ5 Does using isotonic regression within MSTs lead to better predictive performance than parametric alternatives like logistic regression in bid landscape forecasting?

Key findings

• The Isotonic Regression Tree (IRT) model achieved a 5–29% improvement in overall mean squared error (MSE) and 2–14% improvement in AUC over the DSP’s baseline approach across three ad exchanges.
• IRTs outperformed the isotonic regression benchmark (IR) by 7–13% in MSE and 7–15% in AUC, and surpassed IRKM by 1–7% in MSE and 0.6–5% in AUC, demonstrating the value of supervised segmentation.
• With only 10 market segments, IRTs matched or exceeded the performance of IRKM models using up to 1,000 segments, highlighting the method’s efficiency in segment count.
• CMTs reliably identified accurate market segmentations on synthetic and real-world datasets, overcoming model misspecification and showing robustness to overfitting.
• On the Swissmetro dataset, CMTs achieved strong predictive performance compared to common and state-of-the-art benchmarks while providing interpretable user segment definitions.
• The open-source MST implementation demonstrated computational tractability on large datasets, with asymptotic complexity linear in tree depth under mild conditions.