Predictive Modeling for Availability of Inventory

7 Oct 2024

Problem

An allocation engine picks the Target store that ships an online order. If an employee cannot find a unit on the shelf, the order is marked Inventory-Not-Found (INF). Each INF forces a costly re-run of the allocation engine, can split the order into extra packages and delays delivery. The team needs a classifier that scores INF probability per item and store, so the engine avoids risky stores. A false INF rejects a good nearby store, and a missed INF triggers reallocation. Over 90% of samples are non-INF, and all features overlap heavily between classes.

Example from the article: an item that has not sold at a store for a long time is no longer replenished, so an order allocated there often ends as INF.

Solution in brief

Split the data into three segments by business rules, which the post says lowers class imbalance in each. Clean features to cut class overlap, train tree-based boosting on recent plus same-period-last-year data, and retrain around events.

Questions this case answers
  • How to train a classifier for a rare, costly-to-miss event when the positive class is under 10% of samples and its feature distributions overlap with the majority class
  • How to keep a high-volume classifier stable under seasonal and event-driven data drift without continuous manual tuning
  • How to predict whether an item will be missing on the shelf when a store picks an online order?
  • How to train a classifier when over 90% of samples are one class and the feature distributions of both classes overlap?

Tools and methods

  • Business-rule data segmentation — three lower-imbalance segments
  • IQR outlier treatment, feature engineering — less skew and overlap
  • Open-source feature selection — keeps features that differ between classes
  • Tree-based boosting (CatBoost, LightGBM, XGBoost) — the classifier family; the choice is not named
  • Optuna — hyperparameter tuning (in the references)
  • Out-of-time validation — test on a period after training
Continue reading — it's free

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