均值平移聚类

Mean shift treats the observations as samples from an underlying density and searches for its modes, or local peaks. For each seed location, it examines points within a bandwidth and shifts the seed toward a weighted mean of those neighbors. Repeating the update moves the seed uphill on the estimated density surface. Seeds that converge near the same mode are grouped as one cluster. The method can infer the number of clusters from the modes rather than asking for a fixed k. Bandwidth is the key scale parameter. A narrow bandwidth preserves fine local bumps and may create many small modes. A broad bandwidth smooths the density more strongly and can merge nearby peaks, potentially hiding meaningful subgroups. Thus the estimated cluster count is not parameter-free even though k is not supplied. Feature scaling and the kernel distance also influence which points contribute to each update. In a hypothetical two-dimensional dataset with three clear density peaks, seeds placed across the space may move toward those peaks and converge. If bandwidth becomes too large, two neighboring peaks can blend into one; if too small, one peak can fragment into several. Initialization or seed selection affects computational cost and which basins of attraction are explored. Mean shift can be expensive for large datasets because many candidate seeds repeatedly query neighbors. Mean shift is most suitable when density modes are a meaningful definition of groups. It can struggle with varying cluster densities, high-dimensional distance behavior and broad flat regions without clear peaks. It does not provide a calibrated probability of membership. New-point rules vary by implementation; scikit-learn assigns new points to the closest fitted center, a hard-label rule rather than a calibrated membership probability. Inspect the density scale and sensitivity, compare with alternative methods, and assess whether the modes support the downstream task. A cluster is a mode under a chosen kernel and bandwidth, not automatically a natural category.