A recommendation system built on hundreds of thousands of receipts
Retail and e-commerce
Industry and type of organization
A retail chain with its own e-commerce channel and hundreds of thousands of transactions. The receipts held knowledge no one was reading: what customers buy together.
Business problem
The chain had the full purchase history, but that knowledge lay untouched in the receipts. Decisions about what to place next to what, what to suggest at the checkout and what to keep in stock were made from experience, even though the answer was in the receipts themselves: what actually lands together in the basket.
Approach and scope of work
We analyzed hundreds of thousands of receipts with market basket analysis, one of the classic techniques of data mining, extracting association rules: which products go into the basket together and how often. From these patterns we built a recommendation system that suggests combinations backed by sales data.
The same result can be used in several places: in data analysis for inventory planning, in sales as a prompt for staff, and as product recommendations on the website.
The foundation in this project was market basket analysis, which works on anonymous receipts because it is only interested in the basket contents themselves. It covered the whole of sales: both online and the till receipts from brick-and-mortar stores. Collaborative filtering, in turn, is a general recommendation method based on the behavior of customers with similar baskets ("customers like you also bought"). It requires linking purchases to a customer, so it fits wherever that is possible, both online and with a loyalty card in a brick-and-mortar store.
Business effect
hundreds of thousands
receipts analyzed
patterns of products co-occurring
3 uses
cross-selling, inventory, online prompts
one model, many points of use
Product combinations rest on what really happens in the baskets, not on guesswork. The same analysis suggests what to offer at the point of sale, what to place side by side and what to keep in stock.
The patterns from the basket deliver two things that are hard to catch by eye. Slow-moving goods often go in tandem with a specific bestseller, so adding them to such a combination moves stock that would otherwise sit still. You also see non-obvious links between products already in the range that no one had deliberately paired before, and each one is a ready idea for a new sales set.
Scale of the implementation
The analysis covered the chain's full sales history. The longer the history, the more reliable the co-occurrence patterns.
If you have a sales history that no one draws conclusions from, describe your case to us. We will check what patterns can be read out of it.
Related articles in the cm-opti knowledge base
Service area: Data analysis and recommendation systems.