A Survey On Unsupervised Evaluation Criteria For Image Clustering Validation

Abstract

The evaluation of clustering results is the most difficult and frustrating part of cluster analysis. The challenge is to validate the obtained results without any apriori information. Validity indexes are widely used approach for evaluation of clustering results. These approaches can use three criteria: i) external (also called supervised) criteria: this type is based on comparing the obtained results with a previously known result (frequently called ground truth) and compute the similarity, ii) internal criteria (also called unsupervised) criteria: estimate the quality of the result using internal information of the data alone, and iii) relative criteria: this
means multiple usages of one of the two above types ofdifferent results and see which is better than the other. Therefore we can say: depending on the information available and the problem type, different types of indexes might be used for cluster validation. Sometimes due to the complexity of the datasets, one validity index is not sufficient to evaluate the quality of the obtained results, and then a combination of two or more index should be used. In this paper, a basic general review on evaluation criteria is first given and then the focus is spotted on unsupervised criteria as they are much more useful, thanks to their objective functionality.