7. Conclusion

In this part, we discussed the fundamental partitional clustering algorithms and its applications. Partitional clustering produces a partition using K clusters but do not use hierarchical structures. This type of clustering process uses criterion function in a range of error such as mean square error, sum of squared error, and so on. For this reason, partitional clustering is an iterative process.

Some search techniques which include evolutionary process, provide seeking the global or local optima. Search techniques introduce more parameters than K-Means clustering process. There is no theoretical approval for selecting of the best approach. All the given methods can give the best results for different data.

The partitional clustering algorithms need high computational requirements. High computational requirements means that some complex algorithms need advanced operations. These requirements limit their applications in large data. Overcoming this situation, high speed and high-capacity memory-sized computers can be used for clustering processes.

Another method for overcoming high computational requirements is the genetic optimization. For large datasets, genetic optimization should be used with fuzzy clustering. In this technique, all individuals or pixels are genetically optimized so as to fuzzy sets.

Clusters cannot be always sharply separated in some datasets. For this type of data sets, fuzzybased clustering overcomes this crisp clustering. Fuzzy clustering allows to find the nearest optimum cluster to assign into the clusters. This technique provides more information about the data structure.

The most important point of the search techniques of the partitional clustering is the optimum parameter selection. Parameter selection is an optimization problem. Overcoming this optimization problem, parameter selection can be done by using genetic algorithms. Genetic algorithms can be useful solution for very-large scale data sets.

In partitional clustering, determination of cluster size is important. This selection differs from the data sets to data sets. If data set includes more features to classify in a cluster, more clusters will be needed. This cluster number unfortunately is not known for many clustering problems. Generally experiences give the cluster number. Estimation of the cluster number is one of the major problems for validation.

As a result, by using partitional clustering techniques, image understanding can be done by using the given techniques. For many applications such as biomedical image understanding, robotic applications and security applications, these techniques can be useful for pre-processing of some algorithms.
