Online Bayesian learning for e-commerce query reformulation
Search engine performance is usually good on head (high-frequency) queries due to the rich availability of historical behavioral signals on these queries. An over-reliance on past behavioral signals can potentially impact the performance on the tail (low-frequency) queries, where there is a lack of behavioral data. One way to address this issue is to reformulate a tail query into an appropriate head query that the search engine is attuned to and which also preserves the purchase intent of the tail query. From the search engine’s perspective, two queries should be considered equivalent if they lead to the purchase of the same or similar set of products. This property can be used to define a similarity metric over the space of queries, which can then be used to learn a representation for queries using a deep neural network (DNN). However, this similarity metric is noisy for tail queries because a) tail queries are rare, and b) the performance of the search engine for these queries is poor. We address this issue by using Bayesian contextual bandit techniques which refine the representations of the tail queries from the DNN without severely affecting the user experience.