Abstract
Search engine marketing (SEM) is an important channel for the success of e-commerce. With the increasing scale of catalog items, designing an efficient modern industrial-level bidding system usually requires overcoming the following hurdles: 1. the relevant bidding features are of high sparsity, preventing an accurate prediction of the performances of many ads. 2. the large volume of bidding requests induces a significant computation burden to offline and online serving. In this article, we introduce an end-to-end structure of a multi-objective bidding system for search engine marketing for Walmart e-commerce, which successfully handles tens of millions of bids each day. The system deals with multiple business demands by constructing an optimization model targeting a mixture of metrics. Moreover, the system extracts the vector representations of ads via the Transformer model. It leverages their geometric relation to building collaborative bidding predictions via clustering to address performance features' sparsity issues. We provide theoretical and numerical analyzes to discuss how we find the proposed system as a production-efficient solution.