K-Cosine-Medoids Clustering Algorithm

Clustering is a data mining method that aims to partition data into multiple clusters by minimizing inter-cluster similarity and maximizing intra-cluster similarity. K-medoids is one of the most simple and well-known clustering algorithms. However, due to the presence of local outliers and randomly selecting initial medoids, the performance of the algorithm sometimes deteriorates. In this paper, we propose K-cosine-medoids algorithm that extends the idea of K-medoids algorithm and works on minimizing the aforementioned shortcomings. By implementing a systematic approach of selecting initial medoids similar to K-means++ algorithm and using cosine similarity for assigning data points to different clusters and updating medoids, we have observed a significant improvement in terms of accuracy compared to the standard K-medoids algorithm and a number of its variants.