Adaptive random forests for evolving data stream classification

Abstract

Random forests is currently one of the most used machine learning algorithms in the non-streaming (batch) setting. This preference is attributable to its high learning performance and low demands with respect to input preparation and hyper-parameter tuning. However, in the challenging context of evolving data streams, there is no random forests algorithm that can be considered state-of-the-art in comparison to bagging and boosting based algorithms.

The incremental Fourier classifier: Leveraging the discrete Fourier transform for classifying high speed data streams

“…Interestingly, the accuracy ranks (shown in parentheses) of three out of four of these classifiers indicate that they collectively accounted for winners across the entire range of datasets. The accuracy profiles of LB and ARF reinforce results from previous studies where they excelled in performance Bifet et al (2010b), Gomes et al (2017). The strong performance of ARF is to be expected as the basic Random Forest algorithm has proved its superior performance in numerous studies and ARF builds on this success by adapting the forest to concept drift in a data stream environment.…”

“…The nine meta learning algorithms derived from MOA are: Adaptive Random Forest (ARF) Gomes et al (2017), OzaBagASHT (AS), OzaBagADWIN (OB) Bifet et al (2009), LeveragingBag (LB) Bifet et al (2010b), LimAttClassifier (LA) Bifet et al (2010a), AccuracyWeightedEnsemble (AWE) Wang et al (2003), AccuracyUpdatedEnsemble (AUE) Brzeziński & Stefanowski (2011), Anticipative Dynamic Adaptation to Concept Change (AD) Jaber et al (2013a) and Dynamic Adaptation to Concept Changes(DA) Jaber et al (2013b). (9) 49.3(10) 63.7(9) 78.5(8) 73.4(12) 65.8(5) 10 AUE 76.4 (7) 71.1(7) 72.5 (7) 64.2(7) 82.8 (7) 81.8(10) 65.1 (7) (2) 82.9(6) 85.4(7) 69.5(2) 3…”

“…Overall, the ARF classifier had a higher throughput than LB even without the use of a large multi core machine for which it was designed. It was reported in Gomes et al (2017) that its speed increased by a factor of around 3 with the use of a machine that supported a high degree of parallelism with 40 cores. Figures 8 and 9 show the contrast between the IFC and LB classifiers.…”

A Machine Learning Based Approach for Detecting Non-Deterministic Tests and Its Analysis in Mobile Application Testing

“…This preference is attributable to its high learning performance and low demands with respect to input preparation and hyper-parameter tuning. [17] The Random forest algorithm is an ensemble learning method that takes average results of several decision trees to classify its samples. As the name denotes, each decision tree is trained on a random training data subset, perhaps using random features as well.…”

Learning TSK Fuzzy Rules from Data Streams

“…Moreover, a linear function on a corresponding subspace is specified for each leaf of the induced tree, and learning these functions is accomplished using stochastic gradient descent. An ensemble version of FIMTDD (adaptive random forest, ARF-Reg) was proposed by Gomes et al (2018), using an online version of bagging for creating the ensemble members (Oza and Russell, 2001).…”