Product recommendations when a store has little data

1The problem

A store with 300 products and 400 orders a month has very little to learn from. Most of its customers buy once, most products have been bought a handful of times, and a large share of its catalog is new each season. The recommender methods that dominate the literature were built for the opposite case. Item-to-item collaborative filtering was designed for a catalog and customer base large enough that similar items could be computed from purchase overlap alone¹, and matrix factorization learns a hidden vector for every customer and every item from many interactions per customer².

Applied too early, these methods return the same few best sellers to everyone, or they return noise. The question for a small store is not which model is most accurate in general. It is which method its data can support today, and what evidence says it is time to move on.

2Four stages, matched to the data

We propose that a store move through four stages, each one adding a signal the previous stage did not need. A store can sit at different stages in different parts of its catalog: a long-running product line may support item-to-item similarity while new arrivals still rely on rules.

Stage	Signals it needs	What it supports
1. Catalog rules	Product type, collection, tags, price band, stock level	More from the same collection, accessories chosen by the merchant, never an item that is out of stock
2. Co-purchase counts	Orders that contain two or more items	Frequently bought together on the product page and in the cart
3. Item-to-item similarity	Repeated purchases of the same items across many customers, plus product views	Customers who bought this also bought, across the whole catalog
4. Matrix factorization	Returning customers with several interactions each	A personal ranking on the home page, in email and in search

Source: our synthesis of^{1, 2, 3}.

Stage 1 is not a placeholder. Rules written by the merchant encode knowledge no model has yet: which charger fits which device, which items are sold as a set. Stage 2 is the first learned signal and costs almost nothing to compute: count how often two products appear in the same order and show the pairs that recur. It needs no model, it is easy to explain to the merchant, and it fails gracefully, because a pair seen once is simply not shown.

Stage 3 normalizes those counts so that a product bought by everyone is not similar to everything. The item-to-item approach computes the similar-items table offline, which keeps the online cost independent of the number of customers¹. Stage 4 is worth its cost only when enough customers return often enough for a per-customer vector to be estimated; until then, a per-customer model mostly memorizes the overall best sellers.

3Cold start is the normal state

For a large marketplace, a new item or a new visitor is an edge case. For a small store, it is most of the traffic. The cold-start literature treats this directly. Combining content attributes with collaborative data lets a system recommend items that have no purchase history yet, and an evaluation should test that case on its own rather than averaging it into overall accuracy³. Information known about a new visitor, such as the page they arrived on, the market they shop from or what they have in the cart, can stand in for a purchase history until one exists⁴.

In practice this means stage 1 never goes away. When a stage 3 or stage 4 model has nothing to say about an item or a visitor, the store should fall back to the stage below it, not to a global best seller list.

4Judge by carts, not by offline accuracy

A model that predicts held-out purchases better is not necessarily one that sells more. A review of how the business value of recommenders is measured found that studies report it through very different measures, from click-through and adoption to sales and revenue, and that gains in offline accuracy do not reliably translate into gains in those business measures⁵. Part of the reason is that offline accuracy rewards recommending what the customer would have bought anyway.

A small store should therefore move to the next stage only when a controlled comparison shows that the new stage changes what lands in the cart: the share of orders that include a recommended item, the number of items per order, and the revenue from items that were recommended and not already in the cart. If a test at the store's volume cannot detect a difference, that is itself an answer: the simpler stage stays.

5Keep it honest in production

Once a model is live, two things change under it. Shoppers change: seasons, promotions and new markets alter what people buy together. This is concept drift, a change in the relation between inputs and the target over time, and the standard responses are to detect it, to train on a moving window, or to weight recent data more heavily⁶. For a small store the moving window is usually enough, provided it is long enough to hold a meaningful number of orders.

The pipeline also changes: a catalog import renames products, a sync stops, a field goes empty. A rubric for production readiness of machine learning systems lists tests across data, model, infrastructure and monitoring, including checks that input features keep their expected ranges and that a model is not silently stale⁷. The small-store version of that rubric is short: alert when the recommendation table has not been rebuilt on schedule, never recommend an archived or out-of-stock product, and compare each new table with the last one before it goes live.

6Limitations

This is a position paper. The stage boundaries are qualitative; we do not give order counts at which each stage starts to win, because they depend on catalog size, repeat purchase rate and how concentrated sales are. The studies we cite come mostly from large retailers and media services, and their findings may not hold at small volumes. We do not cover search ranking, pricing or content generation, which share some of the same data problems.

7Conclusion

A small store does not need a smaller version of a large retailer's recommender. It needs the method its data supports, a fallback for the many items and visitors with no history, a business measure to decide when to move up, and a few production checks. Rules and co-purchase counts, kept clean, are a defensible place to start.