AI summary by TwinMind
StatQuest: K-means clustering
StatQuest with Josh Starmer
8m 30s
K-Means Clustering Introduction
- Tutorial by Josh Stormer covering k-means clustering fundamentals
- Topics: clustering samples on line, XY graph, heat map data
- Focus on selecting optimal K value
K-Means Algorithm Steps (1D Example)
- Step 1: Select number of clusters (K) - example uses K=3 for three tumor/cell types
- Step 2: Randomly select 3 distinct data points as initial clusters
- Step 3: Measure distance from first point to all 3 initial clusters
- Distance to blue cluster
- Distance to green cluster
- Distance to orange cluster
- Step 4: Assign point to nearest cluster (blue in example)
- Step 5: Calculate mean of each cluster, repeat clustering process
- Continue until clustering doesn't change between iterations
Algorithm Quality Assessment
- Assess quality by summing variation within each cluster
- K-means tracks clusters and total variance
- Algorithm runs multiple times with different starting points
- Returns clustering with lowest total variation
Optimal K Selection Methods
- K=1: Worst case scenario, highest total variation
- K=2: Better, lower total variation than K=1
- K=3: Even better, further reduced variation
- K=4: Continues reducing variation but diminishing returns
- Elbow Plot Method:
- Plot reduction in variance per K value
- Huge reduction at K=3, slower reduction after
- Select K at "elbow" point where improvement levels off
2D Data Clustering (XY Graph)
- Same process as 1D but uses Euclidean distance in 2 dimensions
- Euclidean distance = Pythagorean theorem
- Pick 3 random points, assign to nearest cluster
- Calculate cluster centers, re-cluster until convergence
Multi-Dimensional Data (Heat Maps)
- 2 samples: Rename as x,y coordinates, plot in XY graph
- Distance calculations:
- 2D: √(x² + y²)
- 3D: √(x² + y² + z²)
- 4D: √(x² + y² + z² + a²)
- Pattern continues for higher dimensions
- No plotting required - only distance calculations needed
K-Means vs Hierarchical Clustering
- K-means: Puts data into specified number of clusters (K)
- Hierarchical: Shows pairwise similarity without predetermined cluster count
