Forecasting online user activeness for behavioural targeting: The effect of data sampling

Probability models have been built to model online ad click and conversion, but few studies have examined user activeness, which is the start of any further online behaviour. Using a discrete-time setting, this study builds a three-parameter Bayesian model to forecast user activeness. Users with the same arrival count in the training period are grouped into a segment and their activeness in the test period is forecasted accordingly. The forecasting results are affected by data sparsity and history, while the first factor impacts how to sample the users and the second decides how much historic data should be used in forecasting. Using data from a major ecommerce website, we find that the model performs well when the training period is short while the users are active.