[Paper Review] A Multi-Objective Learning to re-Rank Approach to Optimize Online Marketplaces for Multiple Stakeholders

Phong Nguyen, John Dines+1 more|arXiv (Cornell University)|Aug 2, 2017

Advanced Bandit Algorithms ResearchDecision Sciences1 references16 citations

TL;DR

This paper proposes a novel learning-to-re-rank framework that optimizes online marketplaces for multiple stakeholders—consumers, suppliers, and the intermediary—by re-ranking consumer-focused recommendations to improve both customer conversion and marketplace revenue. Using a kernelized Kendall tau metric for regularization, the method achieves a 50%+ reward rate over a business-rules baseline on margin while maintaining or improving customer relevance.

ABSTRACT

Multi-objective recommender systems address the difficult task of recommending items that are relevant to multiple, possibly conflicting, criteria. However these systems are most often designed to address the objective of one single stakeholder, typically, in online commerce, the consumers whose input and purchasing decisions ultimately determine the success of the recommendation systems. In this work, we address the multi-objective, multi-stakeholder, recommendation problem involving one or more objective(s) per stakeholder. In addition to the consumer stakeholder, we also consider two other stakeholders; the suppliers who provide the goods and services for sale and the intermediary who is responsible for helping connect consumers to suppliers via its recommendation algorithms. We analyze the multi-objective, multi-stakeholder, problem from the point of view of the online marketplace intermediary whose objective is to maximize its commission through its recommender system. We define a multi-objective problem relating all our three stakeholders which we solve with a novel learning-to-re-rank approach that makes use of a novel regularization function based on the Kendall tau correlation metric and its kernel version; given an initial ranking of item recommendations built for the consumer, we aim to re-rank it such that the new ranking is also optimized for the secondary objectives while staying close to the initial ranking. We evaluate our approach on a real-world dataset of hotel recommendations provided by Expedia where we show the effectiveness of our approach against a business-rules oriented baseline model.

Motivation & Objective

• To address the multi-stakeholder recommendation problem in online marketplaces where conflicting objectives of consumers, suppliers, and the intermediary must be balanced.
• To develop a learning-to-re-rank framework that re-ranks consumer-optimized recommendations to also serve secondary objectives, such as intermediary profit and supplier performance.
• To introduce a novel regularization function based on the kernelized Kendall tau metric to preserve relevance while optimizing for secondary objectives.
• To evaluate the method on real-world hotel recommendation data from Expedia, demonstrating improved trade-offs between customer conversion and marketplace revenue.
• To provide a machine learning alternative to ad hoc business rules that currently degrade performance across stakeholders.

Proposed method

• The method frames the problem as a learning-to-re-rank task, starting from an initial ranking optimized for consumer relevance (e.g., clicks and bookings).
• It introduces a novel regularization term based on the kernelized Kendall tau metric to ensure the re-ranked output remains close to the original consumer-optimized ranking.
• The optimization problem is formulated as a linear combination of stakeholder objectives (e.g., customer CVR and intermediary margin), with weights learned via gradient descent.
• The approach uses matrix factorization as the underlying recommendation model and excludes hotel price from features to avoid data leakage into the margin objective.
• A line-search baseline (LS) model is used for comparison, which also uses sigmoid-weighted aggregation of objectives but without the kernelized Kendall tau regularization.
• Evaluation uses NDCG@10 and a risk-sensitive metric measuring the percentage of queries where the LRR method outperforms the baseline on each objective.

Experimental results

Research questions

• RQ1 Can a learning-to-re-rank framework effectively balance consumer relevance and intermediary profitability in online marketplaces?
• RQ2 How does a kernelized Kendall tau-based regularization improve the trade-off between relevance and secondary objectives compared to heuristic business rules?
• RQ3 To what extent does the proposed method outperform a state-of-the-art baseline model in both customer conversion and marketplace revenue?
• RQ4 Does the method maintain consumer relevance while significantly improving intermediary margin across diverse search queries?
• RQ5 Can the approach be generalized to include additional stakeholder objectives such as supplier availability and price competitiveness?

Key findings

• The proposed learning-to-re-rank (LRR) method achieved a 16.7% lift in NDCG@10 for the intermediary margin objective compared to the original consumer-optimized ranking.
• The LRR method incurred only a -5.9% decrease in NDCG@10 for customer preferences (clicks+bookings), outperforming the baseline in terms of risk-reward balance.
• In the risk-sensitive metric, the LRR method achieved a reward rate of over 50% on the margin objective, indicating it outperformed the baseline on more than half of the queries.
• The LRR method demonstrated a systematically lower risk (i.e., fewer losses) than reward on both objectives, showing consistent improvement over the baseline.
• The method achieved a favorable trade-off between customer conversion and marketplace revenue, suggesting it could increase overall platform profitability without sacrificing user experience.
• The results indicate that the kernelized Kendall tau regularization effectively preserves relevance while enabling optimization for secondary objectives, outperforming a rigorously A/B-tested baseline.