An Improved Visual Assessment with Data-Dependent Kernel for Stream Clustering

The advances of 5G and the Internet of Things enable more devices and sensors to be interconnected. Unlike traditional data, the large amount of data generated from various sensors and devices requires real-time analysis. The data objects in a stream will change over time and only have a single access. Thus, traditional methods no longer meet the needs of fast exploratory data analysis for continuously generated data. Cluster tendency assessment is an effective method to determine the number of potential clusters. Recently, there are methods based on Visual Assessment of cluster Tendency (VAT) proposed for visualising cluster structures in streaming data using cluster heat maps. However, those heat maps rely on Euclidean distance that does not consider the data distribution characteristics. Consequently, it would be difficult to separate adjacent clusters of varied densities. In this paper, we discuss this issue for the latest inc-siVAT method, and propose to use a data-dependent kernel method to overcome it for clustering streaming data. Extensive evaluation on 7 large synthetic and real-world datasets shows the superiority of kernel-based inc-siVAT over 4 recently published state-of-the-art online and offline clustering algorithms.