Gradual Forgetting for Adaptation to Concept Drift

The paper presents a method for gradual forgetting, which is applied for learning drifting concepts. The approach suggests the introduction of a time-based forgetting function, which makes the last observations more significant for the learning algorithms than the old ones. The importance of examples decreases with time. Namely, the forgetting function provides each training example with a weight, according its appearance over time. The used learning algorithms are modified to be able to deal with weighted examples. Experiments are conducted with the STAGGER problem using NBC and ID3 algorithms. The results provide evidences that the utilization of gradual forgetting is able to improve the predictive accuracy on drifting concepts. The method was also implemented for a recommender system, which learns about user from observations. The results from experiments with this application show that the method is able to improve the system’s adaptability to drifting user’s interest.