A New Gradual Forgetting Approach for Mining Data Stream with Concept Drift
Yingrong Li, Yang Wei, Anastasiya Kolesnikova, Won Don Lee · 2008
In the real world concepts are often not stable but change with time. The underlying data distribution may change as well. The model built on old data will be necessarily updated. This problem is known as concept drift. Mining concept drifts is one of the most important fields in mining data stream. The paper presents a totally new gradual forgetting approach for mining concept-drift data stream. We firstly utilize UChoo to mine data stream with concept drift. UChoo defines a weight for each instance. The latest data which represents new data distribution has gradually higher weight than old data when time passing. The experiment result shows that the new method performs higher accuracy.