A Comparative Study of Centroid and Medoid based Categorical Data Clustering Methods for Solving Cold-start Recommendation Problem

One efficient solution to solve the cold-start recommendation problem is by exploiting the user demographic information using a clustering method. As the user demographic information contains categorical data, the choice of the clustering method to be used must naturally suitable to the particular data characteristic. There are two popular heuristic categorical data clustering algorithms, i.e., centroid and medoid based. This paper conducts a comparative study towards the implementation of K-Modes of the centroid-based method and K-Approximate Modal Haplotype (K-AMH) of the medoid-based method for solving the cold-start recommendation problem. The experiment results on the MovieLens dataset show that K-AMH achieves the average performance increase of 0.51% in terms of Precision and 0.40% in terms of Normalized Discounted Cumulative Gain (NDCG) to K-Modes. Yet, K-Modes is more lenient to use due to its scalability.