Multi-objective ranking to boost navigational suggestions in eCommerce AutoComplete
Query AutoComplete (QAC) helps customers complete their search queries quickly by suggesting completed queries. QAC on eCommerce sites usually employ Learning to Rank (LTR) approaches based on customer behaviour signals such as clicks and conversion rates to optimize business metrics. However, they do not exclusively optimize for the quality of suggested queries which results in lack of navigational suggestions like product categories and attributes, e.g., "sports shoes" and "white shoes" for query "shoes". We propose to improve the quality of query suggestions by introducing navigational suggestions without impacting the business metrics. For this purpose, we augment the customer behaviour (CB) based objective with Query-Quality (QQ) objective and assemble them with trainable mixture weights to define multi-objective optimization function. We propose to optimize this multi-objective function by implementing ALMO algorithm to obtain a model robust against any mixture weight. We show that this formulation improves query relevance on an eCommerce QAC dataset by at least 13% over the baseline Deep Pairwise LTR (DeepPLTR) with minimal impact on MRR and results in a lift of 0.26% in GMV in an online A/B test. We also evaluated our approach on public search logs datasets and got improvement in query relevance by using query coherence as QQ objective.