A Classifier Graph Based Recurring Concept Detection and Prediction Approach.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Real-world data streams often exhibit recurring concepts, a type of concept drift not fully addressed by current algorithms.
• Existing methods may fail to leverage previously learned information when concepts reappear, leading to suboptimal performance.

Purpose of the Study:

• To propose a novel paradigm for capturing and exploiting recurring concepts in data streams.
• To enhance machine learning performance by effectively handling concept drift and recurring concepts.

Main Methods:

• Incorporation of a distribution-based change detector for concept drift detection.
• Utilizing a classifier graph to store and manage recurring concepts.
• Enabling the reuse of previously learned models through recurring drift detection.

Main Results:

• The proposed approach demonstrates superior performance compared to state-of-the-art algorithms.
• Significant improvements were observed, particularly in scenarios with reappearing concepts.
• Experiments on both synthetic and real-world data streams validated the effectiveness of the method.

Conclusions:

• The novel paradigm effectively captures and exploits recurring concepts in data streams.
• Reusing previously learned models enhances overall learning performance.
• This method offers a significant advancement in handling concept drift with recurring patterns.