Every clustering algorithm covered so far — k-means, spectral clustering — requires you to decide the number of clusters k before running the algorithm. Sometimes that's a reasonable ask. Often it isn't: you genuinely don't know how many natural groups exist in a dataset, and guessing wrong biases the result before you've even started.

Mean Shift clustering doesn't ask for k at all. Instead, it treats the dataset as samples drawn from some underlying probability density, and it finds clusters by locating the modes — the local peaks — of that density. The number of modes found is the number of clusters, discovered automatically from the data's actual shape.

The algorithm: climbing the density hill

Mean Shift works by placing a window (defined by a bandwidth parameter) around every data point and iteratively sliding that window toward the region of highest density nearby — literally walking uphill on the estimated density surface.

1. Start a window centered at each data point (or a representative subsample).
2. Compute the mean of all points falling within the window (typically weighted by a kernel, most commonly Gaussian, so nearer points count more).
3. Shift the window's center to that mean.
4. Repeat steps 2-3 until the window stops moving — it has reached a local density peak (a mode).
5. Group points whose windows converged to the same (or a very close) mode into one cluster.

Every point effectively 'rolls uphill' through the density landscape until it settles at a peak, and points that settle at the same peak belong to the same cluster.

python

Note

Mathematically, mean shift is performing gradient ascent on a kernel density estimate (KDE) of the data — the mean-shift vector at any point is proportional to the gradient of the KDE at that location. This is why the bandwidth parameter matters so much: it's the same bandwidth you'd use to smooth a KDE. A small bandwidth creates a bumpy density surface with many small peaks (over-segmentation); a large bandwidth smooths everything into one or two broad peaks (under-segmentation).

Contrast with the k-means visualization

In the K-Means visualization, you choose k, then watch centroids move toward the mean of their *currently assigned* points — the number of clusters never changes during the run. Mean shift inverts this: there's no fixed set of centroids to begin with, no pre-assigned k, and no discrete assignment step at all. Every point independently climbs its local density gradient, and the final cluster count simply falls out of however many distinct peaks the data happens to have. Picture the same blob dataset from the k-means demo — mean shift would arrive at the same clusters without ever being told there were three of them.

📊 K-Means Clustering Simulator

Interactive

k = 3

Points: 0Iteration: 0Inertia: N/A

K-means vs. mean shift

Property	K-Means	Mean Shift
Number of clusters	Must be specified as k	Discovered automatically from density peaks
Key hyperparameter	k	Bandwidth (window radius)
Cluster shape assumption	Convex, spherical	Arbitrary shape, follows density contours
Sensitivity to outliers	High — pulls centroids off-center	Lower — low-density regions form no peak
Computational cost	Fast, O(n · k · iterations)	Slower, each point iterates its own hill-climb

Note

Mean shift trades 'pick k' for 'pick bandwidth' — it doesn't eliminate the tuning problem, it relocates it. A poorly chosen bandwidth produces either hundreds of tiny spurious clusters or one giant blob. It's also computationally heavier than k-means: each point requires its own iterative hill-climb, and naive implementations are O(n²) per iteration, which is why mean shift is typically used on datasets in the thousands, not millions, of points — classic applications include image segmentation on individual images rather than web-scale clustering.

What's next

Mean shift finds clusters by density, which naturally leaves sparse, low-density points unassigned to any strong peak — a useful signal in its own right. Local Outlier Factor builds directly on the idea of local density to detect anomalies: points sitting in a region far sparser than their neighbors.

I build these systems professionally.

Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.

Work With Me →

Need production AI or MarTech expertise? I build custom LLM workflows, server-side GTM tagging, and analytics infrastructures professionally.

Need custom AI or MarTech setup? Let's build together.

Let's talk →