Stable item codes help preserve search quality

When an e-commerce search system updates the map that links the item-code sequences its model generates to products, can it keep adding items without disrupting retrieval paths learned for existing products? The study tests how retrieval changes with remapping and whether preserving old codes helps.

The authors empirically test identifier updates for e-commerce retrieval and find that preserving old item codes helps protect search quality as new items are added.

The authors used e-commerce search logs to test whether changing existing item identifiers harms retrieval, and whether preserving them helps as new items arrive.

For each query, the system generated identifier candidates and the current map resolved them; HitRate counted queries with the target among the top B resolved candidates.

When both mapping policies were warm-started from the same model on the same accumulated data, preserving old identifiers produced higher beam-size-1 HitRate for old and new items than full remapping. With identifiers frozen, warm-start training also raised HitRate@1000 for both groups compared with no retraining.

The findings support stable old identifiers alongside training for new items in the tested update cycle.

These findings come from offline evaluations and one online test in a specific e-commerce setting. The authors tested a simple warm-start procedure; they caution that a stronger continual-learning method could recover some performance after identifiers change.