dbscan

dbscan

Density-based clustering algorithms.

DBSCAN

library(dbscan)

# DBSCAN clustering
db <- dbscan(data, eps = 0.5, minPts = 5)

# Results
db$cluster  # 0 = noise
db$eps
db$minPts

# Plot
plot(data, col = db$cluster + 1L)
hullplot(data, db)

Finding eps

# k-nearest neighbor distances
kNNdist(data, k = 5)

# Plot to find elbow
kNNdistplot(data, k = 5)
abline(h = 0.5, col = "red")

OPTICS

# OPTICS ordering
opt <- optics(data, eps = 10, minPts = 5)

# Reachability plot
plot(opt)

# Extract clusters
db <- extractDBSCAN(opt, eps_cl = 0.5)
db <- extractXi(opt, xi = 0.05)

# Plot
hullplot(data, db)

HDBSCAN

# Hierarchical DBSCAN
hdb <- hdbscan(data, minPts = 5)

# Results
hdb$cluster
hdb$membership_prob
hdb$outlier_scores

# Plot
plot(hdb)
plot(hdb, show_flat = TRUE)

LOF (Local Outlier Factor)

# Compute LOF scores
lof_scores <- lof(data, minPts = 5)

# Higher scores = more outlier-like
plot(data, cex = lof_scores)

k-NN

# k-nearest neighbors
nn <- kNN(data, k = 5)

# Results
nn$id    # Neighbor indices
nn$dist  # Distances

# Shared nearest neighbors
snn <- sNN(data, k = 5)

Framing

# Points in eps-neighborhood
frNN(data, eps = 0.5)

Predict

# Predict cluster for new points
predict(db, newdata = new_data, data = data)

With Large Data

# Use index for speed
db <- dbscan(data, eps = 0.5, minPts = 5,
  search = "kdtree")  # or "linear", "dist"