Mean Shift Clustering

Mean shift clustering locates the modes of the estimated data density by repeatedly moving points toward the mean of their neighbors within a bandwidth, and assigns cases to the mode they converge to.

평균이동 군집은 대역폭 안 이웃의 평균 쪽으로 점을 반복해 옮겨 추정 밀도의 최빈점을 찾고, 같은 최빈점으로 모이는 사례를 한 군집으로 묶습니다.

I do not know how many clusters my data have; can a method find the density peaks for me?

군집이 몇 개인지 모르는데, 밀도가 높은 봉우리를 알아서 찾아 주는 방법이 있나요?

Mean shift moves each seed point to the mean of observations within a kernel bandwidth until it stops at a local maximum of the density, and cases sharing a mode form a cluster. The number of clusters follows from the bandwidth rather than being fixed in advance, so the bandwidth choice must be reported.

When to use Mean Shift Clustering

• The number of clusters is unknown.
• Clusters correspond to density peaks and may differ in size.
• The number of variables is small to moderate.

When to use something else

• You need explicit noise handling for scattered points; use HDBSCAN instead.
• The data are high-dimensional, where density estimates are unreliable; use k-means after dimension reduction instead.

Assumptions and requirements

• Variables are continuous and on comparable scales.
• A single bandwidth suits all clusters.
• Low to moderate dimensionality.

What the analysis reports

• Number of clusters
• Bandwidth and how it was chosen
• Cluster centers (modes) and sizes
• Silhouette coefficient

How to report it

Mean shift clustering with a flat kernel (bandwidth = [value]; N = [n]) identified [k] clusters (silhouette = [value]).

Related methods

• DBSCAN Density Based Clustering
• HDBSCAN Clustering
• K-Means Clustering

Run Mean Shift Clustering on Quorant