Concept Drift

Traditional approaches to data mining are based on an assumption that the process that generated or is generating a data stream is static. Although this assumption holds for many applications, it does not hold for many others. Consider systems that build models for identifying important e-mail. Through interaction with and feedback from a user, such a system might determine that particular e-mail addresses and certain words of the subject are useful for predicting the importance of email. However, when the user or the persons sending email start other projects or take on additional responsibilities, what constitutes important e-mail will change. That is, the concept of important e-mail will change or drift. Such a system must be able to adapt its model or concept description in response to this change. Coping with or tracking concept drift is important for other applications, such as market-basket analysis, intrusion detection, and intelligent user interfaces, to name a few.