Data stream prediction becomes challenging when faced with issues like concept drift, varying processing time, and memory constraints. Concept drift, defined as changes in data distribution over time, poses a significant hurdle to maintaining the accuracy of prediction systems. To address this, we introduce a method for handling concept drift through a Domain Adaptation approach (CDDA) in a data stream.
The proposed CDDA method passively addresses concept drift by leveraging domain adaptation approaches with multiple sources, simultaneously reducing model execution time and memory consumption. We present two variants of CDDA designed to transfer information from multi-source windows to the target window:
-
Weighted Multi-Source CDDA: A variant focusing on weighted transfer of information from multiple sources.
-
Multi-Source Feature Alignment CDDA: A variant emphasizing feature alignment across multiple sources.
Theoretical insights into the behavior of CDDA are provided, including the derivation of the generalization bound for CDDA in the context of data stream prediction problems.
Extensive experiments have been conducted on both synthetic and real-world data streams to validate the proposed approach's effectiveness. The results demonstrate the robustness and excellent performance of CDDA in handling concept drift.
To run the code, please first install the scikit-multiflow library.
You can check all the evaluation measures by changing the parameters in the main.py file.
- Concept Drift Handling: A Domain Adaptation Perspective
- Authors: Mahmood Karimian, Hamid Beigy
- Journal: Expert Systems with Applications
- Volume: 224
- Pages: 119946
- Year: 2023
- Publisher: Elsevier
- DOI: Concept Drift Handling: A Domain Adaptation Perspective
This project is licensed under the [MIT].