Abstract: Many business applications such as customer modeling or churn prediction require 
data mining among complex and dynamic objects. Customer data are inherently complex; they 
encompass demographics, transactions and product descriptions. They are also dynamic; 
transactions constitute obviously a stream, but the customers themselves, their preferences 
and their needs also change with time. The multi-relational data mining paradigm deals with 
the challenge of model learning over a set of correlated database tables, but is confined to 
static data. In this talk, we discuss model learning over a "multi-table stream", such as a 
stream of customers with their adjoint streams of transactions and products. Model learning 
for such dynamic and complex data requires solutions for (a) the management of the correlated 
streams and their transformation into a single stream for mining and for (b) the impact of 
growing objects upon the model. The first issue involves synchronizing streams that reference 
each other and arrive at different speeds, judiciously forgetting objects that reference each 
other. The second issue involves monitoring how an object "grows" as other objects reference it, 
and then deciding which objects to involve in model learning – depending, among others, on object 
size. We report on experiments for model learning on bank customer data. 
[Siddiqui & Spiliopoulou,  SSDBM'09] [Siddiqui & Spiliopoulou, DS'09] [Siddiqui & Spiliopoulou, SSDBM'10]