Drift Detection Using Stream Volatility
Current methods in data streams that detect concept drifts in the underlying distribution of data look at the distribution difference using statistical measures based on mean and variance. Existing methods are unable to proactively approximate the probabi
- PDF / 1,117,651 Bytes
- 16 Pages / 439.37 x 666.142 pts Page_size
- 24 Downloads / 232 Views
Department of Computer Science, University of Auckland, Auckland, New Zealand {dtjh,ykoh,gill}@cs.auckland.ac.nz 2 Huawei Noah’s Ark Lab, Hong Kong, China [email protected]
Abstract. Current methods in data streams that detect concept drifts in the underlying distribution of data look at the distribution difference using statistical measures based on mean and variance. Existing methods are unable to proactively approximate the probability of a concept drift occurring and predict future drift points. We extend the current drift detection design by proposing the use of historical drift trends to estimate the probability of expecting a drift at different points across the stream, which we term the expected drift probability. We offer empirical evidence that applying our expected drift probability with the state-ofthe-art drift detector, ADWIN, we can improve the detection performance of ADWIN by significantly reducing the false positive rate. To the best of our knowledge, this is the first work that investigates this idea. We also show that our overall concept can be easily incorporated back onto incremental classifiers such as VFDT and demonstrate that the performance of the classifier is further improved. Keywords: Data stream
1
· Drift detection · Stream volatility
Introduction
Mining data that change over time from fast changing data streams has become a core research problem. Drift detection discovers important distribution changes from labeled classification streams and many drift detectors have been proposed [1,5,8,10]. A drift is signaled when the monitored classification error deviates from its usual value past a certain detection threshold, calculated from a statistical upper bound [6] or a significance technique [9]. The current drift detectors monitor only some form of mean and variance of the classification errors and these errors are used as the only basis for signaling drifts. Currently the detectors do not consider any previous trends in data or drift behaviors. Our proposal incorporates previous drift trends to extend and improve the current drift detection process. In practice there are many scenarios such as traffic prediction where incorporating previous data trends can improve the accuracy of the prediction process. For example, consider a user using Google Map at home to obtain a fastest route to a specific location. The fastest route given by the system will be based on c Springer International Publishing Switzerland 2015 A. Appice et al. (Eds.): ECML PKDD 2015, Part I, LNAI 9284, pp. 417–432, 2015. DOI: 10.1007/978-3-319-23528-8 26
418
D.T.J. Huang et al.
how congested the roads are at the current time (prior to leaving home) but is unable to adapt to situations like upcoming peak hour traffic. The user could be directed to take the main road that is not congested at the time of look up, but may later become congested due to peak hour traffic when the user is en route. In this example, combining data such as traffic trends throughout the day can help arrive at a better prediction. Similarly, using
Data Loading...
