- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper proposes Mass-Aware Spectral-Clustering (MASC), a three-step clustering algorithm. The method rely on the notion of "supernode", which are groups of nodes that are formed using a single-linkage hierarchy. These supernodes are then ordered in a tree strucutre on which spectral clustering is applied. On the theory side, the autors show that this reduced problem is exactly a RatioCut on a sample-level graph, and they give a Davis-Kahan bound on the stability of the embedding. Experimentally, MASC reaches the best ACC and NMI on the showcased dataset.
Critical strengths
Overall I like the simplicity of the appraoch. Once the supernodes are built, the graph to cut is just a tree, which is cheap to cluster. The method is light on hyper-parameters, besides the number of clusters , the only parameter is , and it seems that there is no kernel bandwidth or neighbourhood size to tune, which is often a painful part in traditional spectral clustering. Moreover, the method is well-motivated, and bridging density-based and spectral clustering is an interesting and challenging problem.
Critical weaknesses
I would say that there are still a few weaknesses in the paper that would need a treatment.
- The paper says is searched over , but I couldn't find how the reported value is chosen. If it is the value that maximises ACC against the ground-truth labels, this should be stated carefully, as this could be an important factor in the method's performance in the experimental result. Fig 4a is reassuring on this aspect as it would seem that the selection of is stable, but this is shown only for five datasets. Results at a fixed on all datasets would make the comparison more convincing, alternatively, having a way to select without using the labels would also be interesting to have.
- The contribution of each stage doesn't seem so clear to me. The supernodes come from CHC, but CHC itself doesn't appear in the comparison table. Without it, it is hard to tell how much of the performance comes from the tree and the spectral step, and how much is already achieved by the proper hierarchy. Related methods such as FINCH or DBSCAN would also be natural baselines.
- One of the main weakness of the paper to me is the clarity. I found the paper hard to follow. The text relies a lot on abstract phrasing and new vocabulary, sometimes before it is properly defined. For example: "The saliency , the product of mass and separation by which DPC selects centers, serves here as evidence." A figure showing the full pipeline would help a lot. The notations could also be lighter. The same edge weight appears as , , and . The level index in is only introduced implicitly, is used both for the dimension and for the dissimilarity and and introduce confusion.
Questions for authors
- How is the reported selected? Does the selection use the ground-truth labels?
- The peak-affinity graph is a tree, so a -way partition into connected parts amounts to deleting edges. How does the spectral relaxation compare with simpler tree cuts, for instance removing the edges with the largest saliency, or solving the objective exactly when is small?
- What is the value of in each datasets ? How does it compare to ? How well does CHC alone perform?