Using Deep Exponential Families as Generative Models in Marketing Data Fusion

Nowadays more and more data is collected from a broad variety of sources. Often in marketing applications the combined information of multiple sources is of interest. Combining these different sources of information is a process called data fusion. Currently, data fusion literature in marketing applications is limited and consists of simple models that have shortcomings in their performance. This work uses deep exponential families (DEF) as a generative model for the task of data fusion in a marketing application. The DEF allows to maintain correlational structures without relying on so-called statistical matching. Using both a simulation study and a private survey data set by Nielsen, it is demonstrated that the DEF strongly outperforms existing data fusion models. As such it can be clearly concluded that DEFs form a valuable instrument in the quantitative marketing literature.