The clustering analysis is a subject that has been interesting researchers from several areas, such as health (medical diagnosis, clustering of proteins and genes), marketing (market analysis and image segmentation), information management (clustering of web pages). The clustering algorithms are usually applied in Data Mining, allowing the identification of natural groups for a given data set. The use of different clustering methods for the same data set can produce different groups. So, several studies have been led to validate the resulting clusters. There has been an increasing interest on how to determine a consensus clustering that combines the different individual clusterings, reflecting the main structure in clusters inherent to each of them, as a perspective to get a higher quality clustering. As several techniques of consensus clustering have been researched, the present work focuses on problem of finding the best partition in the consensus clustering. We analyze the most referred techniques in literature, the consensus clustering techniques with different mechanisms to achieve the consensus, i.e.; Voting mechanisms; Co-association matrix; Mutual Information and hyper-graphs; and a multi-objective consensus clustering existing on literature. In this paper we discuss these approaches and a comparative study is presented, that considers a set of experiments using two-dimensional synthetic data sets with different characteristics, as number of clusters, their cardinality, shape, homogeneity and separability, and a real-world data set based on hand's biometrics shape, in context of people parental recognition. With this data we intend to investigate the ability of the consensus clustering algorithms in correctly cluster a child and her/his parents. This has an enormous business potential leading to a great economic value, since that with this technology a website can match data, as hand's photographs, and say if A and B are related somehow. We conclude that, in some cases, the multi-objective technique proved to outperform the other techniques, and unlike the other techniques, is little influenced by poor clustering even in situations like noise introduction and clusters with different homogeneity or overlapped. Furthermore, shows that can capture the performance of the best base clustering and still outperform it. Regarding to real data, no technique was capable of identifying a person's mother/father. However, the research of distances between hands from a person and its father, mother, siblings, can retrieve the probability of that person being his/her familiar. This doesn't enable the identification of relatives but instead, decreases the size of database for seeking the matches.