spectral clustering vs hierarchical clustering

Question

Can anyone please explain that is there any advantage of using hierarchical clustering over spectral clustering? I know how they work but I want to know in which situations its better to use hierarchical clustering over spectral clustering.

Answer 1

Hierarchical Clustering:

• It is good for taxonomical data.
• Keep connecting similar points until you end up with a single cluster containing all the data points.
• From the grouping that you get after combining all the data, it is called a dendrogram.
• Depending on how many clusters you want, you cut the dendrogram at a level.

---

Spectral Clustering:

• It is for clustering data which is in complex shape as K-means and Density-based clustering fail in these cases.
• Treating the data points as vertices of a graph, connect the vertices that are close enough. Hence you also choose an ε value. All vertices closer than ε are connected.
• Unlike Hierarchical clustering, you do not get a fully connected single graph (unless you take a high ε value and there are no non-connected components). Instead you get multiple connected components, each of them representing a cluster.

Answer 2

Hierarchical clustering is usually faster and produces a nice dendrogram to study.

Dendrograms are very useful to understand if you have a good clustering.

Furthermore, hierarchical clustering is very flexible. You can use different distance functions and different linkage strategies.

Spectral clustering has an interesting theory; but to work really well, your data must already consist of well separated "connectedd components". When you have very noisy data, good luck.

Answer 3

• with hierarchical clusters, you have additional information about the relationship btw clusters bc you can slice at a different number of clusters and see how clusters group together at lower k.
• With hierarchical, you can compute clustering once, which depending on dataset size can be very advantageous, and slice the result at different k values
• Agglomerative clustering (a hierarchical method) produces the same result every time, unlike spectral clustering which has a random component.
• Because of the additional constraint of hierarchality, hierarchical methods usually perform worse at any given k compared to non-hierarchical methods by objective metrics like silhouette and inertia