Learning to Rank and Discover for E-Commerce Search

Abstract

E-Commerce (E-Com) search is an emerging problem with multiple new challenges. One of the primary challenges constitutes optimizing it for relevance and revenue and simultaneously maintaining a discovery strategy. The problem requires designing novel strategies to systematically “discover” promising items from the inventory, that have not received sufficient exposure in search results while minimizing the loss of relevance and revenue because of that. To this end, we develop a formal framework for optimizing E-Com search and propose a novel epsilon-explore Learning to Rank (eLTR) paradigm that can be integrated with the traditional learning to rank (LTR) framework to explore new or less exposed items. The key idea is to decompose the ranking function into (1) a function of content-based features, (2) a function of behavioral features, and introduce a parameter epsilon to regulate their relative contributions. We further propose novel algorithms based on eLTR to improve the traditional LTR used in the current E-Com search engines by “forcing” exploration of a fixed number of items while limiting the relevance drop. We also show that eLTR can be considered to be monotonic sub-modular and thus we can design a greedy approximation algorithm with a theoretical guarantee. We conduct experiments with synthetic data and compare eLTR with a baseline random selection and an upper confidence bound (UCB) based exploration strategies. We show that eLTR is an efficient algorithm for such exploration. We expect that the formalization presented in this paper will lead to new research in the area of ranking problems for E-com marketplaces.