Solving Product Substitutions, A Big Problem in Grocery E-Commerce – Through Self-Supervised ML

Speaker: Jawad Ahmed, Staff Applied Scientist, Loblaw Digital

"Background:
Loblaw Companies Ltd is the largest grocery retailer in Canada. It operates multiple popular banners with Real Canadian Superstore, No Frills, and T&T being some of the most popular ones. E-commerce of grocery has become a significant part of the business accounting for more than $2 billion in sales per year.

Problem:
Shopping for groceries online is an inherently different process than shopping in person. We take for granted the in-store shopper’s ability to make quick decisions on the fly when faced with the issue of product availability.

We fulfill from stores to ensure freshness which has a very dynamic inventory. This makes promises of items collected, sometimes a day or two after the order depending on the customer’s delivery date, affected by many factors - some of which we cannot control. Thus, we need a solution to substitute items that are out of stock at the time of picking to make sure the customer experience is minimally impacted. While shopping at a physical store, a customer can make a suitable choice of an alternative. In the e-commerce process of grocery shopping, either the customer has to make a selection of the substitute, or the Loblaw employee picking the order on behalf of the customer needs a relevant suggestion on the best substitute for the given item, personalized for the given customer.

Loblaw has historical data available on what selection was made by customers from the list of various possible substitute options available for a given item. Additionally, there is data available on the choices made by pickers - the employees who shop at the store to fulfill customers’ orders. This provides us an opportunity to tailor product similarities toward product substitutions that are tied to business metrics.

Solution:
We explored multiple solutions to solve this problem. Our most promising solution that we wish to present leverages features extracted from text descriptions and images of products. In this talk, we will discuss how our approach evolved over time and how this cutting-edge self-supervised method is a big improvement over the traditional techniques."