Personalized Recommender System Design for an Online B2B Platform
We collaborated with IndiaMart, the largest online B2B platform in India, to design and implement a personalized recommender system for participant firms. 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. The core of our algorithm involves predictive and prescriptive analytics to first learn sellers’ preferences from their historical behavior using an attribute-based logit model and feature engineering, and then to determine the display order of RFQs on the sellers’ webpages. Our algorithm is novel in two respects: (1) new closeness metrics to learn sellers’ evolving preferences from high-dimensional and sparse data; (2) methods to counter the problem of class imbalance using a new resampling strategy we propose, which we call Panel Data Augmentation Technique (PDATE). We implemented our algorithm in a controlled field experiment at IndiaMart since January 2020, resulting in a significant improvement in the acceptance of recommended RFQs from the top of the webpage. The improvement is consistent and sustained over time including during the pandemic. After a pilot test with 21 sellers, the company has now adopted the algorithm for 2,000 sellers in a gradual expansion to a larger scale.
Team:
Vishal Gaur, Cornell University
Xiaoyan Liu, Cornell University
Gaurang Manchanda, IndiaMART