CNODE: clustering of set-valued non-ordered discrete data

This paper introduces a clustering technique named 'Clustering of set-valued Non-Ordered DiscretE data' (CNODE), in which each data item is a vector having a set of non-ordered discrete values per dimension. Since usual definitions of distance like Euclidean and Manhattan do not hold for 'non-ordered discrete data space' (NDDS), other measures like Hamming distance are often used to define distance between vectors having single-valued discrete dimensions. Such type of distance is not meaningful for set-valued dimensions and hence, we propose a similarity measure based on set intersection for clustering set-valued vectors. We also suggest a new measure for determining quality of clustering named 'lines of clustroids' (LOC) for this type of data. In contrast to other existing clustering techniques in NDDS, CNODE does not rely on any kind of pre-processing of dataset. Experiments with synthetic and real datasets show that CNODE is robust to data variations, scalable to large dataset size and efficient for high dimensions.