Incremental Belief-Peaks Evidential Clustering

Chaoyu Gong, Sihan Wang, Zhi-gang Su

Abstract: Despite evidential clustering has been widely employed in pattern recognition problems characterized by imprecise and uncertain membership, its application in the realm of big data remains constrained by excessive computational complexity and limited computational resources. To bridge this research gap, this paper introduces an Incremental Evidential Clustering (IEC) method based on stream data clustering and belief-peaks, a technique that has demonstrated exceptional effectiveness in detecting cluster centers under numerous cases. Commencing with an initial small dataset, IEC progressively scales its operations to accommodate escalating volumes of data. Through judicious employment of scalar transformations, IEC cost-effectively updates the affinity matrix, ensuring incremental adjustments. Targeting solely the altered segments of the affinity matrix, recalibration of cluster centers takes place currently. For voluminous datasets, users can also adjust the update frequency according to their specific requirements. Compared to state-of-the-art (SoTA) stream clustering algorithms, IEC demonstrates better clustering accuracy and comparable runtime across four benchmark datasets. IEC markedly diminishes runtime contrasted with other SoTA evidential clustering algorithms.