Data Mining and Tree-based Optimization

Consider a large collection of objects, each of which has a large number of attributes of several different sorts. We assume that there are data attributes representing data, attributes which are to be statistically estimated or predicted from these, and attributes which can be controlled or set. A motivating example is to assign a credit score to a credit card prospect indicating the likelihood that the prospect will make credit card payments and then to set a credit limit for each prospect in such a way as to maximize the over-all expected revenue from the entire collection of prospects. In the terminology above, the credit score is called a predictive attribute and the credit limit a control attribute. The methodology we describe in the paper uses data mining to provide more accurate estimates of the predictive attributes and to provide more optimal settings of the control attributes. We briefly describe how to parallelize these computations. We also briefly comment on some of data management issues which arise for these types of problems in practice. We propose using object warehouses to provide low overhead, high performance access to large collections of objects as an underlying foundation for our data mining algorithms.