Meta Learning of the Effective Cost Per Mille

We develop meta features for the prediction of the effective cost per mille (eCPM) and total advertiser value (TAV) metrics commonly used in production systems containing click-through rate (CTR) and conversion rate (CVR) prediction models. eCPM is an overall performance metric based on the revenue incurred by the advertisements offered by the underlying CTR or CVR prediction models and reflects overall revenue efficiency, incorporating traffic, conversion, and bidding data. TAV is an overall performance metric prioritizing advertisers’ benefit. We run experiments on the meta learning task of the eCPM and TAV prediction of CTR and CVR prediction models already built. We engineer 27 features to improve the performance, including derived metrics like normalized CTR and CVR, predicted eCPM, efficiency, and estimated cost progression degree. We evaluate 6 machine learning models on two datasets from separate regional data and achieve state-of-the-art results. Stacking of random forest, M5P, and M5Rules models obtains the best results. Compared to previous approaches, our framework significantly improves RMSE and offers a robust solution for optimizing advertising revenue. We also provide a ranking of the top 15 features.