Spectral clustering in eye‐movement researches

Abstract

Eye tracking is a widely used technology to capture the eye movements of participants completing different tasks. Several eye-tracking parameters are measured, which later can be used to characterize the gazing pattern of the individuals.Clustering based on the path walked on by the participants may enable the researchers to create clusters based on the unconscious personality and thinking style. Common clustering methods generally are unable to handle path data; hence, new dynamic variables are needed. Spectral clustering can handle these types of data well. Spectral clustering handles clustering as a graph partitioning problem without making specific assumptions on the form of the clusters and uses eigenvectors of matrices derived from the data. This way, data are mapped to a low-dimensional space, which can be easily clustered. Different food choice tasks were presented, and each of the 149 participants had to choose 1 product of the presented 4 and later from 8 alternatives. A new measure was introduced based on all 3 consecutive points from the fixations, and the areas of the triangles formed by these 3 points were computed. The new eye-movement index captures the temporal variation and also considers the orientation of the fixation points.Spectral clustering resulted 5 balanced clusters defined by Dunn, Silhouette, and C-indices. Results were compared to the most widely applied hierarchical and centroid-based clustering (k-means) methods. Spectral clustering achieved the best results in clustering indices and cluster sizes proved to be more balanced;hence, it outperforms the commonly used applied hierarchical and k-means.

Comparison of supervised learning statistical methods for classifying commercial beers and identifying patterns

“…The MFA-based expert classification was compared with the most widely applied hierarchical (agglomerative hierarchical clustering [AHC] Euclidean distance and Ward's method), centroid-based (k-means), and the novel spectral clustering methods. 22 Comparison was done by the following cluster validity indices: Dunn, 23 Silhouette, 24 and C-index. 25 The error calculations have been performed by using TANAGRA 26 and all other calculations in R system 27 (for hierarchical clustering: hclust package; for calculating cluster validity indices: clusterCrit package; for performing spectral clustering: SNFtool package).…”

“…In the training error calculation, 0.7 was used as a proportion of the train set (350 instances) and repeated the training 10 times by randomly selecting the new train set. The MFA‐based expert classification was compared with the most widely applied hierarchical (agglomerative hierarchical clustering [AHC] Euclidean distance and Ward's method), centroid‐based ( k ‐means), and the novel spectral clustering methods . Comparison was done by the following cluster validity indices: Dunn, Silhouette, and C‐index .…”