Learning Product Improvement from Consumer Evaluations

Product Improvement (PI) involves making minimal changes to an existing product to enhance its market performance. A key source of information for PI is Consumer Evaluation (CE), provided either in aggregate (e.g., sales volume) or descriptive (e.g., online reviews) form. Existing PI approaches analyze the descriptive CE of the focal product, which limits their applicability when descriptive CE is missing, abstract, or interdependent. More importantly, current methods are primarily descriptive and do not generate actionable PI recommendations, despite their considerable value to practitioners. We address these challenges with a new approach that leverages both aggregate and descriptive CE across the entire market to prescribe specific PI recommendations. We first develop Product Segmentation (PS) trees, a type of decision tree with a customized objective function designed to identify the features that best segment products by market performance. We then build a constrained shortest-path algorithm over the PS tree structure to prescribe minimal improvements for any underperforming product. To enhance robustness, we ensemble multiple PS trees, each trained with distinct objectives, into a novel PS forest. Using both real and synthetic data, our approach achieves 73% average precision in PI and improves upon the best benchmarks by 11%. Lastly, we conduct a conjoint analysis to behaviorally validate that consumers prefer the PI recommendations generated by our method over the original underperforming products. Our theoretical and empirical results show that the proposed approach addresses key challenges in using CE for PI and offers interpretable and actionable decision support for practitioners. This paper was accepted by D. J. Wu, information systems. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2025.00055 .