Mean shift clustering in ML Python - Model Pipeline Trace

Model Pipeline - Mean shift clustering

Mean shift clustering groups data points by shifting them towards the densest area nearby, finding clusters without predefining their number.

Data Flow - 5 Stages

1Input Data

300 rows x 2 columns→Raw 2D points representing data samples→300 rows x 2 columns

[[1.2, 3.4], [2.1, 3.9], [0.9, 3.0], ...]

↓

2Bandwidth Selection

300 rows x 2 columns→Choose radius to define neighborhood size for shifting→Scalar value (bandwidth)

bandwidth = 1.0

↓

3Mean Shift Iteration

300 rows x 2 columns→Shift each point towards mean of neighbors within bandwidth→300 rows x 2 columns

Point [1.2, 3.4] shifts to [1.5, 3.6]

↓

4Convergence Check

300 rows x 2 columns→Repeat shifting until points move less than threshold→300 rows x 2 columns

Final shifted points cluster around centers

↓

5Cluster Assignment

300 rows x 2 columns→Assign points to clusters based on proximity to shifted centers→300 rows x 1 column

[0, 0, 1, 1, 2, 2, ...] cluster labels

Training Trace - Epoch by Epoch

Epochs
5 | *****
4 | ****
3 | ***
2 | **
1 | *
    Loss decreases as points shift less

Epoch	Loss ↓	Accuracy ↑	Observation
1	N/A	N/A	Initial shifting of points begins
2	N/A	N/A	Points move closer to local density peaks
3	N/A	N/A	Shifts become smaller as points near centers
4	N/A	N/A	Most points converge; clusters form clearly
5	N/A	N/A	Convergence reached; final cluster centers stable

Prediction Trace - 5 Layers

Layer 1: Select a single data point

Layer 2: Find neighbors within bandwidth

Layer 3: Calculate mean of neighbors

Layer 4: Shift point to mean

Layer 5: Repeat until convergence

Model Quiz - 3 Questions

What does the bandwidth control in mean shift clustering?

A The initial position of points

B The size of the neighborhood to find nearby points

C The number of clusters to create

D The speed of convergence

Key Insight

Mean shift clustering finds clusters by moving points towards dense regions without needing to know cluster count beforehand. It works well for discovering clusters of any shape by iteratively shifting points until they settle near local peaks.