Mining Evolving Data Streams with Particle Filters. (21st September 2015)

• Record Type: Journal Article
• Authors: Fok, Ricky; An, Aijun; Wang, Xiaogang
• Is Part Of: Computational intelligence. Volume 33:Number 2(2017)
• Page Start: 147
• Page End: 180
• Publication Date: 2015-09-21
• Subjects: concept drift -- ensemble methods -- high dimensional data stream mining; Artificial intelligence -- Periodicals; Computational linguistics -- Periodicals; 006.3
• DOI: 10.1111/coin.12071
• Languages: English
• ISSNs: 0824-7935

Abstract

Abstract : We propose a particle filter‐based learning method, PF‐LR, for learning logistic regression models from evolving data streams. The method inherently handles concept drifts in a data stream and is able to learn an ensemble of logistic regression models with particle filtering. A key feature of PF‐LR is that in its resampling, step particles are sampled from the ones that maximize the classification accuracy on the current data batch. Our experiments show that PF‐LR gives good performance, even with relatively small batch sizes. It reacts to concept drifts quicker than conventional particle filters while being robust to noise. In addition, PF‐LR learns more accurate models and is more computationally efficient than the gradient descent method for learning logistic regression models. Furthermore, we evaluate PF‐LR on both synthetic and real data sets and find that PF‐LR outperforms some other state‐of‐the‐art streaming mining algorithms on most of the data sets tested.