Parallel hierarchical clustering using weighted confidence affinity

Abstract

There have been many attempts for clustering categorical data such as market basket dataset. However, most of categorical clustering approaches belong to partitional clustering which requires at least one input parameter (e.g., the minimum intra-cluster similarity or the desired number of clusters). In this paper, we propose a parallelised hierarchical clustering approach for categorical data (PH-clustering) using vertical data structures. In order to minimise the impact of low support items, we devise a weighted confidence (WC) affinity function to compute the similarity between clusters. Based on our analysis of the major clustering steps, we adopt a partial local and partial global approach to reduce computation time as well as to keep network communication at minimum. Load balance issues are addressed especially during the data partitioning phase. Our experimental results on standardised market basket data show that the proposed weighted confidence affinity measure is more accurate than other contemporary affinity measures in the literature and that our parallel clustering approach provides magnitudes of time improvements over sequential clustering especially over larger data sizes. Our results also indicate that the number of items/attributes in the dataset has a more drastic impact on performance than the number of transactions/tuples.