Title: Online multi-dimensional regression analysis on concept-drifting data streams

Authors: Chandima Hewa Nadungodage; Yuni Xia; Pranav S. Vaidya; Yu Chen; Jaehwan John Lee

Addresses: Department of Computer and Information Science, Indiana University – Purdue University Indianapolis, 723 West Michigan Street, Indianapolis, IN 46202-5132, USA ' Department of Computer and Information Science, Indiana University – Purdue University Indianapolis, 723 West Michigan Street, Indianapolis, IN 46202-5132, USA ' Department of Electrical and Computer Engineering, Indiana University – Purdue University Indianapolis, 723 West Michigan Street, Indianapolis, IN 46202-5132, USA ' Department of Electrical and Computer Engineering, Indiana University – Purdue University Indianapolis, 723 West Michigan Street, Indianapolis, IN 46202-5132, USA ' Department of Electrical and Computer Engineering, Indiana University – Purdue University Indianapolis, 723 West Michigan Street, Indianapolis, IN 46202-5132, USA

Abstract: Stream data is generated continuously in a dynamic environment, with huge volume and fast changing behaviour. In order to perform regression on data streams, it is required to incrementally reconstruct the regression model as new stream data flows in. However, due to the tremendous data volume, it is impossible to scan the entire data stream multiple times to re-compute the regression model parameters. Therefore, one-scan algorithms are required for such streaming applications. In this paper, we investigate online multi-dimensional regression analysis of concept-drifting data streams, and present two algorithms, approximate stream regression (ASR) and ensemble stream regression (ESR). ASR approach dynamically re-computes the regression function parameters, considering not only the data records of the current window, but also a synopsis of the previous data. ESR approach trains an ensemble of regression models from sequential chunks of the data stream, and then computes the weighted average of their predictions. Experiments show that the proposed methods are not only efficient in time and space but also able to generate better fitted regression functions compared to the existing stream regression algorithms such as sliding window regression and incremental stream regression.

Keywords: online regression analysis; multi-dimensional regression analysis; data streams; approximate stream regression; ensemble stream regression; concept drift; data modelling; regression models.

DOI: 10.1504/IJDMMM.2014.065146

International Journal of Data Mining, Modelling and Management, 2014 Vol.6 No.3, pp.217 - 238

Published online: 23 Oct 2014 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article