Automated System for Objectives Driven Merchandising

Important results from the project

The purpose was to make goal optimization tractable in e-commerce stores. It should be easy to change goal of an entire site. As an example, before and during fire sales a store might want to optimize for conversion, and by default, the site should optimize towards either revenue or profit. Functions affected are search, product listings, and recommendations. Our goal is to enable optimization towards a business goal by modifying global settings. We achieved the goal by implementing a business app where it is possible to choose among three goals: conversions, revenue, and profit.

Expected long term effects

KPI-optimization for search and product listings was obtained. For product recommendations, we developed two hybrid variants. One is a bandit algorithm and the other is based on combinatorial optimization. The latter is designed to facilitate KPI-optimization but is not a reinforcement learning algorithm. We have also enabled merchandisers to increase exposure of sets of products across the site. The store can obtain a plurality of goals with this tool. Examples include increased sales at the end of product life cycle, and increased exposure of new products to kick-start sales.

Approach and implementation

Theoretical inquiries proceeded according to project plan. Practical implementation followed agile development procedures. We performed a user survey. The result thereof was aggregated into a POC, that contained a demo of basic functionality, as well as a description of the project as a power point presentation. The project was evaluated iteratively by existing customers. Given the input from the POC, we developed a business app for KPI-optimization, including an AB-test tool for customers. The app was extended to include a function for advanced product exposure management.