Many of the multichannel extracellular recordings of neural activity consist of attempting to sort spikes on the basis of shared characteristics with some feature detection techniques. Then spikes can be sorted into distinct clusters. There are in general two main statistical issues: firstly, spike sorting can result in well-sorted units, but by with no means one can be sure that one is dealing with single units due to the number of neurons adjacent to the recording electrode. Secondly, the waveform dimensionality is reduced in a small subset of discriminating features. This shortening dimension effort was introduced as an aid to visualization and manual clustering, but also to reduce the computational complexity in automatic classification. We introduce a metric based on common neighbourhood to introduce sparsity in the dataset and separate data into more homogeneous subgroups. The approach is particularly well suited for clustering when the individual clusters are elongated (that is nonspherical). In addition it does need not to select the number of clusters, it is very efficient to visualize clusters in a dataset, it is robust to noise, it can handle imbalanced data, and it is fully automatic and deterministic.