Predicting Click-Through Rate Using Keyword Clusters

Click-through rate is a quantity of interest to online advertisers, search engine optimizers, and sponsored search providers alike. The rate at which users click on advertisements presented to them serves as both a metric for evaluating advertising effectiveness and a financial tool for cost and revenue projection. To predict future click-through rate, investigators typically use historical click information since it provides tangible examples of user behavior. In some cases plentiful historical data are available and this method provides a reliable estimate. More often, however, insufficient historical data exists and creative aggregation must fill the gap. In this work, we hypothesize that different terms have an inherently different likelihood of receiving a sponsored click. For example, the search terms “digital camera” and “brain structure” clearly express more and less shoppingoriented intent, respectively. We seek to estimate a termlevel click-through rate (CTR) reflecting these inherent differences. At times even aggregation to the term level leaves us with insufficient historical data for a confident estimate, so we also propose the use of clusters of related terms for less frequent, or even completely novel, terms. We reviewed other broad estimates of CTR (rank CTR and query-volume decile CTR) and compared them to the estimates computed using hierarchical clusters of related terms. We found that using historical data aggregated by cluster leads to more accurate estimates on average of term-level CTR for terms with little or no historical data.