Computational Learning Models For Drifting Concepts

In this paper, we explore computational learning models for learning concepts classes where the target concept changes with time, a scenario known in the literature as concept drift. The importance of this topic is immediately evident in the dynamic nature of real-world concepts; if a phenomenon can be represented by a static concept, then it often does not even require learning. The most obvious examples of the idea of concept drift are in the domain of customer preference: For instance, a teenagers may for a time be interested in the music of Britney Spears, but (if their parents are lucky) as they mature, their tastes might shift more towards Bob Dylan. Advertisers that initially learned the customer’s interest in Britney Spears may perform poorly if their learning algorithm is not able to adjust with the changing tastes of this consumer. Common solutions to this problem are to either discard old data, or prevent any concept from getting so deeply ingrained that the algorithm cannot recover quickly from a concept change. In Littlestone’s analysis of the Weighted Majority Algorithm, he presents a variant that assigns a lower bound on the weight of any particular expert with the goal of making the algorithm more tolerant to concept drift[LW94]. More recently, Mesterharm analyzes a version of Winnow that uses the same idea[M02]. In this paper, we will consider some other approaches to handling concept drift. The subject of concept drift has been approached both from the vantage point of Computational Learning Theory, and from the vantage point of Artificial Intelligence. This paper includes some discussion of the latter perspective, which tends to focus on testing,