Predictive model performance: offline and online evaluations
Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining
| Hi-index | 0.00 |
We study the problem of estimating the value of a contextual ad impression, and based upon which an ad network bids on an exchange. The ad impression opportunity would materialize into revenue only if the ad network wins the impression and a user clicks on the ads, both as a rare event especially in an open exchange for contextual ads. Given a low revenue expectation and the elusive nature of predicting weak-signal click-through rates, the computational cost incurred by bid estimation shall be cautiously justified. We developed and deployed a novel impression valuation model, which is expected to reduce the computational cost by 95% and hence more than double the profit. Our approach is highly economized through a fast implementation of kNN regression that primarily leverages low-dimensional sell-side data (user and publisher). We also address the cold-start problem or the exploration vs. exploitation requirement by Bayesian smoothing using a beta prior, and adapt to the temporal dynamics using an autoregressive model.