Variable Density Based Genetic Clustering

From the existing clustering techniques, spatial density-based ones register one of the most promising results in detecting arbitrary shaped data, being robust to outliers and not restricted by various data distributions. The existing literature contains a plethora of density based algorithms but in all cases one or multiple global parameters need to be set, parameters that are seldom easy to set requiring in depth knowledge about the analyzed data. This paper proposes a parameter-free novel genetic clustering algorithm with an original method for encoding clustering solutions relying on density based clustering parameters. Within each clustering solution genotype, gene position, defined by several density based clustering attributes, plays a key role for recovering the encoded partition. Each gene defined density-based cluster can only attract object not already attracted by previously defined clusters. The proposed encoding scheme allows for always valid crossover results, with great offspring variations even when using simple crossover operators. While not requiring any input parameters, experiments involving multiple clustering validation indices as fitness criteria, across both synthetic and real data sets, show comparable results with existing density-based clustering techniques.