Unreliability of clustering results in sensory studies and a strategy to address the issue

Abstract

Researchers commonly use hierarchical clustering (HC) or k-means (KM) for grouping products, attributes, or consumers. However, the results produced by these approaches can differ widely depending on the specific methods used or the initial “seed” aka “starting cluster centroid” chosen in clustering. Although recommendations for various clustering techniques have been made, the realities are that objects in groups can, and do, change their clusters. That can impact interpretation of the data. Researchers usually does not run the clustering algorithms multiple times to determine stability, nor do they often run multiple methods of clustering although that has been recommended previously. This study applied hierarchical agglomerative clustering (HAC), KM and fuzzy clustering (FC) to a large descriptive sensory data set and compared attribute clusters from the methods, including multiple iterations of same methods. Sensory attributes (objects) shuffled among clusters in varying ways, which could provide different interpretations of the data. That frequency was captured in the KM output and used to form the “best possible” clusters via manual clustering (MC). The HAC and FC results were studied and compared with KM results. Attribute correlation coefficients also were compared with clustering information. Using results from one clustering approach may not be reliable, and results should be confirmed using other clustering approaches. A strategy that combines multiple clustering approaches, including a MC process is suggested to determine consistent clusters in sensory data sets.