Personalized Recommender System Design for IndiaMART’s Online B2B Platform

Note: The Innovative Applications in Analytics Award (IAAA) is a prestigious award developed by the Analytics Society of INFORMS to recognize creative and unique application of a combination of analytical techniques in new areas. Presented each year by the Analytics Society along with Kinaxis and Adelphi University, the award attracts submissions from around the world whose work is judged by a panel of experts. This is the second in a series of brief articles describing the work of the 2021 IAAA finalists.

IndiaMart is the largest B2B platform in India, serving more than 119 million buyers and 6.5 million seller firms and transacts about 464 million requests for quotation each year across a wide range of industries and geographies. The dominant business model of IndiaMart works as follows: Buyers post their business requirements on the platform to invite quotations from sellers; sellers join the platform as subscribing members to discover buyer requests relevant to their businesses; and the platform generates personalized recommendations to sellers to facilitate market clearing and earn subscription revenue.

The core of this task is to accurately predict which requests a seller is likely to accept. Buyer requests are time-sensitive and arrive at a rate of more than 1 million per day. Furthermore, personalized recommendations must be generated in real time within milliseconds when sellers log on to the platform in order to provide a fast and seamless browsing experience. A higher prediction accuracy leads to better quality matching, which drives seller and buyer satisfaction, increased engagement and revenue growth, whereas a lower prediction accuracy can cause goodwill costs and loss of engagement. Thus, IndiaMart requires a scalable, efficient and accurate algorithm to generate recommendations.

Our team has worked together since July 2019 to solve this problem in four major stages: We first created a data pipeline to easily share up-to-date data across the team, then using historical transactions, we created and tested a choice model to accurately predict requests matched to heterogeneous sellers. In the third stage we tested the prescriptive recommendation-generating algorithm for live implementation and conducted a controlled pilot test during the January-July 2020 timeframe, and finally in the fourth stage, we scaled the algorithm to more than 10% of the users on the platform.

Our work has successfully contributed to the platform by achieving significant and sustained improvements in the performance of the recommendation page. In particular, the average rank of accepted requests improved from 37.95 pretreatment to 16.21 during treatment, and the percentage of requests accepted from the top 25 positions improved by 8.2% in the pilot test (see Figure 1). This work was especially beneficial to society during the COVID-19 pandemic when India implemented a complete lockdown in March-June 2020 and supply chains for personal protective equipment (PPE) moved online to the company’s platform. Our algorithm enabled the platform to automatically learn the changes in seller preferences and prioritize personal PPE, hand sanitizer, oxygen concentrators, etc., thus providing a valuable service to fulfill this need.

The methodology developed in our work contributes to online marketplace analytics and is widely applicable to other retailing and marketplace problems that face similar challenges. Our main innovation is the development of a new technique that we call the panel data augmentation technique, or PDATE, to mitigate class imbalance due to an overrepresentation of accepted requests and underrepresentation of rejected requests, which is a standard problem in clickstream data and makes it difficult to reliably train a prediction rule. Our method also required further innovations to learn preferences for heterogeneous sellers from high-dimensional and sparse data.

Reference

1. Gaur, Vishal, Xiaoyan, Liu, 2020, “Personalized Recommendation System Design for an Online B2B Platform,” Nov. 11, https://ssrn.com/abstract=3902710.