Using TB-Sized Data to Understand Multi-Device Advertising

In this study, we combine the conversion funnel theory with statistical model to understand multi-device advertising. We investigate the important question of how the distribution of ads on multiple devices affects the consumer path to purchase. To handle the sheer volume of TB sized impression data, we develop a MapReduce framework to estimate the non-stationary Hidden Markov Model in parallel. To accommodate the iterative nature of the estimation procedure, we leverage the Apache Spark framework on a corporate cloud computing infrastructure. We calibrate the model with hundreds of millions of impressions for 100 advertisers. Our preliminary results show increasing the diversity of device for ads delivery can consistently encourage consumers to become more engaged. In addition, advertiser heterogeneity plays an important role in the variety of the conversion process.