Get access

Clustering algorithm based on mutual K-nearest neighbor relationships



Clustering algorithms for data with varying densities have been investigated in the past and there are some data situations and clustering needs that are not handled well by these algorithms. We present here an algorithm for such situations in which multiple, possibly overlapping, clusters exist and need to be identified by their density as the defining characteristic. In this paper, we define the idea of mutual K-nearest neighbors (MKNN) relationship based on inter-point affinities and use it as a basis for discovering the above-mentioned types of clusters. With a synthetic and two real-world datasets we show that our algorithm delivers the type of clustering results and robustness that we seek to achieve, and the performance is better than what is achievable by other algorithms. Statistical Analysis and Data Mining 2012