Light feed-forward networks for shard selection in large-scale product search

Large-scale information retrieval systems store documents in different shards. Shard selection enables cost-effective retrieval by searching only relevant shards for the query. Most existing shard selection algorithms focus on web search, and rely on text similarity between the query and shard corpora. In contrast, in e-commerce product search, shards are defined according to product categories, and most queries imply product category intent. Such characteristics are yet to be leveraged for shard selection. In this work, we formulate shard selection in product search as a multi-label query intent classification problem. We show that light feed-forward neural networks, with language-independent features, suffice to achieve high performance for this recall-oriented task. The simple architecture allows for low-latency shard selection in the early retrieval process. We evaluate the model in terms of cost reduction and impact on the relevance of retrieved documents, both in offline simulation and online A/B testing. Without degrading customer experience, we achieve double-digit percentage of search engine cost reduction in multiple locales, and the model has been deployed to serve Amazon Search customers worldwide.