Personalized Recommendation System Design for an Online B2B Platform

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 field experiments. In this problem, buyers place requests for quotation (RFQs) to the platform, sellers respond by accepting or declining those RFQs, and the objective of the platform is to design a recommendation system by matching RFQs with sellers based on likelihood of acceptance. Our research is conducted in collaboration with IndiaMart, the dominant online B2B platform in India serving approximately 119 million buyer firms and 6.4 million seller firms in more than 71 million products and services. The main challenges in our problem are short timespan of RFQs, high-dimensional, sparse, and heterogeneous transaction data, and class imbalance such that the volume of `accepted' records is significantly larger than `declined' records.We construct an attribute-based logit classifier to solve this problem, apply feature engineering to address high-dimensionality, and propose a new resampling approach, Panel Data Augmentation Technique, to counter class imbalance. To synthesize the trade-off between canonical predictive metrics and determine the optimal resampling strategy, we propose a novel aggregate measure based on operational costs. Our results demonstrate a high out-of-sample predictive accuracy and a significant gain in the recall for minority class in offline testing. Two controlled field experiments conducted at IndiaMart show that our methods lead to significant improvements in the top-5, top-10, and top-25 contributions from the recommendations page. Our method has been adopted by the platform for all their customers