A Scalable Approach to Fuzzy Rough Nearest Neighbour Classification with Ordered Weighted Averaging Operators

Abstract

Fuzzy rough sets have been successfully applied in classification tasks, in particular in combination with OWA operators. There has been a lot of research into adapting algorithms for use with Big Data through parallelisation, but no concrete strategy exists to design a Big Data fuzzy rough sets based classifier. Existing Big Data approaches use fuzzy rough sets for feature and prototype selection, and have often not involved very large datasets. We fill this gap by presenting the first Big Data extension of an algorithm that uses fuzzy rough sets directly to classify test instances, a distributed implementation of FRNN-OWA in Apache Spark. Through a series of systematic tests involving generated datasets, we demonstrate that it can achieve a speedup effectively equal to the number of computing cores used, meaning that it can scale to arbitrarily large datasets.

Scalable Approximate FRNN-OWA Classification

“…As we noted in previous work [9], in practice it is desirable to work with weight vectors w and w of length k m. Formally, these can be interpreted as full weight vectors where the last or first m − k values are equal to 0. The first benefit is presentational, since this restricts the range of possible OWA operators to precisely those we are interested in: OWA operators that approximate max and min by emphasising the largest and smallest elements, respectively.…”

“…There has been a limited number of attempts to apply fuzzy rough sets in the context of Big Data. For a current overview, see [9] (summarised in Table II). All previous studies only used fuzzy rough sets for feature or prototype selection, and only one tested their implementations on real datasets with more than a million instances [30].…”

“…All previous studies only used fuzzy rough sets for feature or prototype selection, and only one tested their implementations on real datasets with more than a million instances [30]. In [9], we presented an implementation of FRNN-OWA in Apache Spark that could in principle be used to distribute computation time over as many computing nodes as desired. While this is a useful tool, which enabled us to use training sets of more than 10 million instances to classify unseen instances, it requires a considerable amount of computational infrastructure.…”

“…In previous work, we have presented an implementation of FRNN-OWA in Apache Spark that reduces its run time through parallellisation accross a cluster of computers [9]. While this provides a method to perform classification with large datasets, it cannot speed up FRNN-OWA by a factor larger than the number of processor cores available, and it would be more effective if we could reduce the time complexity of FRNN-OWA itself.…”

Fast and Scalable Approaches to Accelerate the Fuzzy k-Nearest Neighbors Classifier for Big Data

“…In the big data environment [14], the kNN and FkNN algorithms have been key to solving different machine learning problems such as fuzzy-rough based NN classification [15], time-series forecasting [16] or data preprocessing to obtain quality data [17]. In this work, we are focused on standard classification.…”