HICCS: Hybrid Instance Clustering Based on Compact Sets

Data clustering is a widely used technique in Pattern Recognition, Artificial Intelligence and Data Mining. It has been shown to be useful in many practical domains such as text mining, bioinformatics, image segmentation and wireless sensor networks. Recently, there has been a growing emphasis on clustering mixed (or hybrid) as well as incomplete datasets. In this paper we introduce HICCS, a hierarchical clustering for hybrid and incomplete data. Our proposal has several novel characteristics, because it uses compact sets as initial groups instead as single objects, it uses the similarity between cluster representatives as inter-cluster dissimilarity, avoiding additional computations and incorporates a merging strategy to evade order-dependencies. By comparing our proposal with other existing algorithms, we argue that HICCS is one of the best available clustering methods for handling mixed and incomplete datasets. We note that HICCS estimates the true partitions of data, having a significant better performance with respect to other algorithms. It can improve the cluster quality and applicability. The effectiveness of HICCS is shown for seven artificial data sets of varying complexities and fifteen real-life data sets with hybrid and incomplete data.