Hellinger Distance Trees for Imbalanced Streams

Abstract

Abstract-Classifiers trained on data sets possessing an imbalanced class distribution are known to exhibit poor generalisation performance. This is known as the imbalanced learning problem. The problem becomes particularly acute when we consider incremental classifiers operating on imbalanced data streams, especially when the learning objective is rare class identification. As accuracy may provide a misleading impression of performance on imbalanced data, existing stream classifiers based on accuracy can suffer poor minority class performance on imbalanced streams, with the result being low minority class recall rates. In this paper we address this deficiency by proposing the use of the Hellinger distance measure, as a very fast decision tree split criterion. We demonstrate that by using Hellinger a statistically significant improvement in recall rates on imbalanced data streams can be achieved, with an acceptable increase in the false positive rate.

Fifty years of pulsar candidate selection: from simple filters to a new principled real-time classification approach

“…Data streams are quasi-infinite sequences of information, which are temporally ordered and indeterminable in size (Gaber et al 2005;Lyon et al 2013Lyon et al , 2014. Data streams are produced by many modern computer systems (Gaber et al 2005) and are likely to arise from the increasing volumes of data output by modern radio telescopes, especially the SKA.…”

“…Data streams are produced by many modern computer systems (Gaber et al 2005) and are likely to arise from the increasing volumes of data output by modern radio telescopes, especially the SKA. However many of the effective supervised machine learning techniques used for candidate selection do not work with streams (Lyon et al 2014). Adapting existing methods for use with streams is challenging, it remains an active goal of data mining research (Yang & Wu 2006;Gaber et al 2007).…”

“…It is designed to maximise classification performance on candidate data streams, which are heavily imbalanced in favour of the non-pulsar class. It is the first candidate selection algorithm designed to mitigate the imbalanced learning problem (He & Garcia 2009;Lyon et al 2013Lyon et al , 2014, known to reduce classification accuracy when one class of examples (i.e. non-pulsar) dominates the other.…”

“…The Gaussian-Hellinger Very Fast Decision Tree (GH-VFDT) is an incremental stream classifier, developed specifically for the candidate selection problem (Lyon et al 2014). It is a tree-based algorithm based on the Very Fast Decision tree (VFDT) developed by Hulten et al (2001).…”

“…When applied to a data stream containing 10,000 non-pulsar candidates for every legitimate pulsar (HTRU data obtained by Thornton (2013)), it raised the recall rate from 30 to 86 per cent (Lyon et al 2014). This was achieved using candidate data described using the features designed by Bates et al (2012) and Thornton (2013).…”