Personalized Recommendation System Design for an Online B2B Platform
39 Pages Posted: 12 Aug 2021
Date Written: November 11, 2020
We formulate the problem of designing a personalized recommendation system for an online business-to-business (B2B) marketplace, propose a method to solve it, and evaluate results using a controlled pilot test. Our research is conducted in collaboration with IndiaMart, the dominant online B2B platform in India serving approximately 60 million buyer firms and 5 million seller firms in more than 50 million products and services. In this problem, buyers place requests for quotation (RFQs) to the platform, and the objective of the platform is to match the RFQs with suitable sellers with the highest likelihood of acceptance. Our problem entails two major challenges: (1) high-dimensional and sparse data regarding product category and spatial engagement, and (2) class imbalance because the volume of "accepted" records in historical clickstream data is significantly larger than that of "declined" records. We propose new variables motivated by the choice estimation literature to address high-dimensionality, and evaluate alternative approaches including the Synthetic Minority Over-sampling Technique (SMOTE) and a new resampling approach, which we call Panel Data Augmentation Technique (PDATE), to counter class imbalance. Our method yields a significant improvement in out-of-sample predictive accuracy. A controlled pilot test conducted at IndiaMart shows that our method provides a consistent and significant improvement in the quality of recommendations sustained over time.
Keywords: Online marketplace, recommendation system, high-dimensionality, class imbalance, field experiment, choice estimation, machine learning
JEL Classification: C51, C52, C55, C93, L81, M11, M21
Suggested Citation: Suggested Citation