CADENCE: Offline Category Constrained and Diverse Query Generation for E-commerce Autosuggest

Abstract

Query AutoComplete (QAC) or AutoSuggest is the first place of user interaction with an e-commerce search engine. It is critical for the QAC system to suggest relevant and well-formed queries for multiple possible user intents. Suggesting only the historical user queries fails in the case of infrequent or new prefixes. Much of the recent works generate synthetic candidates using models trained on user queries and thus have these issues: a) cold start problem as new products in the catalogue fail to get visibility due to lack of representation in user queries b) poor quality of generated candidates due to concept drift and c) low diversity/coverage of attributes such as brand, color & other facets in generated candidates. In this paper, we propose an offline neural query generation framework - CADENCE - to address these challenges by a) using both user queries and noisy product titles to train two separate neural language models using self-attention memory networks, b) adding category constraints during the training and query generation process to prevent concept drift c) implementing customized dynamic beam search to generate more diverse candidates for a given prefix. Besides solving for cold start and rare/unseen prefix coverage, CADENCE also increases the coverage of the existing query prefixes through a higher number of relevant and diverse query suggestions. We generated ∼700K new offline queries, which have resulted in significant improvement in recall, reduction in product cold start, and increased coverage of attributes. Online A/B tests also show a significant impact on QAC usage, downstream search click-through rates, and product conversion.