Returned together - Recommendation pairs cleaned with return data

A furniture retailer’s “goes well with” slot on product pages ran on association rules mined from past orders. For each pair of items, two numbers were computed: support, the share of all orders containing both items, and lift, how much more often the two appear together than chance alone would predict. Pairs with enough support and a high lift became recommendations. It was cheap and explainable, and click-through on the slot was healthy.

In furniture, returns are very expensive: bulky items travel both ways, and often cannot be resold as new. So the company started an effort to reduce them, beginning with what the returned orders had in common. One pattern stood out: some of the most popular pairs were also among the most returned. A dining table and a particular chair sold together constantly. In the products photo the wood looked like a perfect match, but in real life the two shades were clearly different, and customers ordered both, then sent back one or both of them.

The rules could not see this. They only knew that two items landed in the same basket more often than chance. An order that was kept counted the same as one where something went back a week later. The slot was recommending combinations that looked good on screen and disappointed on arrival, yet each counted as a success in the click and sales reports.

The first instinct was to build something clever: a model that predicts whether a pair will be liked. It sounded reasonable, but it would need weeks of work, and nobody could explain to a merchandiser why a given pair had vanished.

The simpler fix used data the company already had. For each pair, two counts were kept over a rolling window: how many orders contained both items, and how many of those orders saw at least one of the two returned. The share of joint orders with a return is the pair’s return rate:

$$\text{pair return rate} = \frac{\text{orders where either item was returned}}{\text{orders containing both items}}$$

Pairs whose rate sat above a cut-off were dropped from the candidate list before ranking. Everything else stayed the same. The change was a filter, not a new ranker.

Two details mattered. First, a return of either item counts. Sending back just one of the two is already a bad sign: the add-on was bought because the slot suggested it, or the main item was bought in the expectation that the pair would work, and in both cases the recommendation did not deliver. Second, pairs with very few joint orders were left alone, because a rate based on a handful of orders swings wildly.

The effect showed within a couple of weeks. Click-through on the slot dipped slightly, but the share of recommended pairs that customers actually kept went up, and fewer returns followed a click on the slot. A slot that looked less impressive in the click report was doing its job better.