Empirical Comparison of K-Means Initialization Methods for Card Sorting Data Analysis

This is the authors' abstract. We don't add key points for this paper.

Card sorting is a user research method used to gain insights into how people categorize and comprehend information. Analysis of the collected data typically involves using clustering algorithms, such as the K-means algorithm. However, it is well established that the efficiency and accuracy of the K-means is greatly affected by the initialization method used. This paper empirically compares seven K-means initialization methods for analyzing card sorting datasets using six metrics of clustering quality. We used a card sorting dataset from a real-world study for redesigning the information architecture of a University e-learning platform, involving 112 participants and 50 cards. It was found that the Best Merge Method provides the best results when analyzing open card sort datasets from 30 or fewer participants. When the number of participants exceeds 30, the Best Merge Method and Actual Agreement Method are the most effective approaches to employ.