Feature engineering for the current Out of stock detection ML model, Trax Retail (Retail Watch team)

Michael Ben Mergui

Data Science Fellows February 2021 Cohort



In this project, the focus was on feature engineering and feature selection to increase the current model in
production performance.
Features explored to improve the Metamodel performance:
1) New `distance from empty` feature based on pixel distance from a product which is an out-of-stock
candidate on a shelf in the supermarket aisle. This part included feature engineering, binning, trimming,
running multiple options of engineering of the new feature in the model and seeing how the model
2) Analyze the difference between current and past neighbors. Objective was to see if there is a difference in
neighbors and see if it is linked to the product being in stock or out of stock. Using an API called the prior API
created by another Trax team.
3) Looking at D-2 and D-14 days prior information and the likelihood of the out of stock product to be in its
most likely prior location. Using an API called the prior API created by another Trax team.


1. It is difficult to accept that a feature that you research on for 2 weeks does not give good model results
but that is normal with research in ML, sometimes it does not work!

2. At first it has been difficult on the first week to quickly get to know the databases, the
data, and how to access them per project.


1. I assessed the first feature as more performant on all metrics when it was replaced by older features
used in the current model in production. This was a great success for the team and the new feature
was included in the 4 features used by the Metamodel for its predictions. The feature is going to be put
in production shortly.
2. I identified multiple other new features to research on and explore that could potentially improve the
performance of the Metamodel.
Future project development
On my project I researched the features 2 and 3 and did not manage to get to any conclusion. However, I
identified many further ways to approach and explore these 2 new features and identified other ones to
A future project for a next intern could be to work on these new features in a similar internship to mine.

Share this post

Share on facebook
Share on twitter
Share on linkedin
Share on email