Using Proximity Graph Cut for Fast and Robust Instance-Based Classification in Large Datasets. (29th November 2021)

• Record Type: Journal Article
• Authors: Protasov, Stanislav; Khan, Adil Mehmood
• Other Names: Cheong Siew Ann Academic Editor.
• Is Part Of: Complexity. Volume 2021(2021)
• Publication Date: 2021-11-29
• Subjects: Chaotic behavior in systems -- Periodicals; Complexity (Philosophy) -- Periodicals; 003
• DOI: 10.1155/2021/2011738
• Languages: English
• ISSNs: 1076-2787

Abstract

Abstract : K-nearest neighbours (kNN) is a very popular instance-based classifier due to its simplicity and good empirical performance. However, large-scale datasets are a big problem for building fast and compact neighbourhood-based classifiers. This work presents the design and implementation of a classification algorithm with index data structures, which would allow us to build fast and scalable solutions for large multidimensional datasets. We propose a novel approach that uses navigable small-world (NSW) proximity graph representation of large-scale datasets. Our approach shows 2–4 times classification speedup for both average and 99th percentile time with asymptotically close classification accuracy compared to the 1-NN method. We observe two orders of magnitude better classification time in cases when method uses swap memory. We show that NSW graph used in our method outperforms other proximity graphs in classification accuracy. Our results suggest that the algorithm can be used in large-scale applications for fast and robust classification, especially when the search index is already constructed for the data.