An effective multiobjective approach for hard partitional clustering

Clustering is an unsupervised classification method in the field of data mining. Many population based evolutionary and swarm intelligence optimization methods are proposed to optimize clustering solutions globally based on a single selected objective function which lead to produce a single best solution. In this sense, optimized solution is biased towards a single objective, hence it is not equally well to the data set having clusters of different geometrical properties. Thus, clustering having multiple objectives should be naturally optimized through multiobjective optimization methods for capturing different properties of the data set. To achieve this clustering goal, many multiobjective population based optimization methods, e.g., multiobjective genetic algorithm, mutiobjective particle swarm optimization (MOPSO), are proposed to obtain diverse tradeoff solutions in the pareto-front. As single directional diversity mechanism in particle swarm optimization converges prematurely to local optima, this paper presents a two-stage diversity mechanism in MOPSO to improve its exploratory capabilities by incorporating crossover operator of the genetic algorithm. External archive is used to store non-dominated solutions, which is further utilized to find one best solution having highest F-measure value at the end of the run. Two conceptually orthogonal internal measures SSE and connectedness are used to estimate the clustering quality. Results demonstrate effectiveness of the proposed method over its competitors MOPSO, non-dominated sorting genetic algorithm, and multiobjective artificial bee colony on seven real data sets from UCI machine learning repository.