A dynamic similarity weighted evolving fuzzy system for concept drift of data streams

Abstract

Financial markets and weather prediction are generating streaming data at a rapid rate. The frequent concept drifts in these data streams pose significant challenges to learners during training and prediction. Concept drift means the distribution of data in the data stream changes dramatically at any time, and it can lead to decreased effectiveness in many data-driven systems. So, obtaining online models that adapt to concept drift is essential. In the face of concept drifts in data streams, most proposed evolving fuzzy systems (EFSs) suffer from two problems. First, it is difficult to quickly adapt to a drastic change in a short period, such as sudden drift. Second, most ensemble EFSs adjust the weights according to the error, which will easily lead to the risk of under-training and repeated training of the base model. To solve the above problems, we propose a new type of ensemble EFS called dynamic similarity weighted evolving fuzzy system (DSW-EFS). Unlike existing ensemble EFSs, DSW-EFS assigns weights to base learners based on the similarity between data distributions, and each base learner describes only one data distribution in the data stream. This ensemble method enables DSW-EFS to describe multiple data distributions in the data stream and quickly adapt to multiple concept drifts. To ensure the accuracy of drift detection, we propose a Gaussian mixture model (GMM)-based concept drift detection algorithm that can obtain the similarity between data distributions. This detection method can achieve high accuracy based on a smaller sliding window size. The DSW-EFS model is tested from concept drifts and prediction accuracy in the experimental part. The results show that DSW-EFS can quickly adapt to concept drift by assigning weights to base learners based on similarity when concept drifts occur. Moreover, DWS-EFS achieves high prediction accuracy on most datasets.