Machine Learning and Assortment Management in Stores

Mikhail Shchekalyov27 May 2023

Problem

Yandex Eda delivers groceries from partner stores. Each store sends a feed with products and their stock, but the stock does not always match the shelf. The storefront shows products that are missing or scarcer than displayed. The picker then has to find a substitute and agree on it with the customer. This lowers service quality and retention, slows delivery, ties up pickers and couriers, and can cancel the whole order.

Example from the article: the storefront shows a product as in stock, the picker cannot find it, and the product is removed from the cart or replaced with an alternative.

Solution in brief

Predict for every product in a store whether it will be found, and hide the ones predicted as unavailable. A CatBoost classifier trained on order logs, with numeric popularity features for the long tail, goes beyond a reactive "hide after a failure" heuristic. A small global control group, where nothing is hidden, supplies training data free of selection bias.

Questions this case answers
  • How to predict whether a listed item will actually be available at order time when the supplier's stock data is unreliable and the model's own filtering removes its training signal
  • How to keep an unbiased stream of labels for a classifier that hides items, using a global control group, when rare categories have very few samples
  • How to predict whether a product is really on the shelf and hide unavailable items from the storefront
  • How to handle the long tail of rarely ordered products and categories in a retail classifier

Tools and methods

  • Reactive heuristic — hide a product for N hours after a "not found" signal; baseline.
  • Binary classification with CatBoost — predicts whether a product will be found.
  • Numeric proxies for popularity — average stock, average stock delta and frequency of stock zeroing replace part of the categorical features.
  • Global control group in an A/B test — no model, nothing hidden; source of unbiased training data and online metrics.
Continue reading — it's free

Google or a one-time email code. No password.