Conventional stream mining methods assume that each data instance is seen only once and is forgotten after being processed. Consider for example a classifier that distinguishes between normal network accesses and attacks. This classifier reads each data instance (access operation) once and must adapt to new types of attack. However, the data to be analyzed in many business applications are not simple instances, but complex, nested objects that contain streams of data instances. Customer data are such an example: they encompass some stationary information, as well as transactions like purchases, service requests, product reviews etc. To learn and maintain customer segments, a company needs learning methods that derive and adapt models upon the complex objects and the streams feeding them.

In IMPRINT we distinguish between perennial objects, which contain data instances, and the stream of data instances themselves. The challenges of mining perennial objects are manifold. They include learning upon objects that grow as new transactions arrive, the comparison of objects that differ in size and age, and their efficient maintenance. In IMPRINT, we will design, develop and evaluate adaptive learning methods that deal with the above challenges.

Publications