Interpreting Advertiser Intent in Sponsored Search

  • Bhanu Vattikonda ,
  • Santhosh Kodipaka ,
  • Hongyan Zhou ,
  • Vacha Dave ,
  • ,
  • Alex C. Scoeren

Proceedings of the 21st ACM Conference on Knowledge Discovery and Data Mining (KDD) |

Published by ACM

Search engines derive revenue by displaying sponsored results along with organic results in response to user queries. In general, search engines run a per-query, on-line auction amongst interested advertisers to select sponsored results to display. In doing so, they must carefully balance the revenue derived from sponsored results against potential degradation in user experience due to less-relevant results. Hence, major search engines attempt to analyze the relevance of potential sponsored results to the user’s query using supervised learning algorithms. Past work has employed a bag-of-words approach using features extracted from both the query and potential sponsored result to train the ranker. We show that using features that capture the advertiser’s intent can significantly improve the performance of relevance ranking. In particular, we consider the ad keyword the advertiser submits as part of the auction process as a direct expression of intent. We leverage the search engine itself to interpret the ad keyword by submitting the ad keyword as an independent query and incorporating the results as features when determining the relevance of the advertiser’s sponsored result to the user’s original query. We achieve 43.2% improvement in precision-recall AUC over the best previously published baseline and 2.7% improvement in the production system of a large search engine.