|
Submitted by
Assigned_Reviewer_4
Q1: Comments to author(s).
First provide a summary of the paper, and then address the following
criteria: Quality, clarity, originality and significance. (For detailed
reviewing guidelines, see
http://nips.cc/PaperInformation/ReviewerInstructions)
The authors present a theoretical analysis of spectral
clustering in the degree corrected stochastic block model. Following
earlier work of Chaudhuri et. al. they analyze spectral clustering with
the regularized normalized Laplacian, first showing that the population
Laplacian reflects the correct block structure, and then showing that the
"noisy" graph Laplacian has a similar spectrum. Finally, they bound the
number of mis-clustered nodes. They also consider an additional step of
normalizing rows of the spectral embedding and show that the resulting
bounds are related to the statistical leverage scores.
Overall the
paper is well written (a few typos are pointed out at the end), addresses
an interesting problem and provides a much simpler (albeit weaker)
analysis as compared to the earlier work of Chaudhuri et. al.
Here
are some issues that the authors should clarify:
1. I don't quite
understand why the passage before equation (6) is an appropriate
definition of mis-clustered nodes. There is already a natural definition
from the ground truth clustering. It would help if the authors clarified
the relation between the k-means 0/1 error and the set \mathcal{U}. Why
are the nodes in \mathcal{U} the only ones on which k-means (which does
not have access to population centroids, say with random initialization)
makes errors? If not, it would help avoid confusion to not use the term
"mis-clustering rate".
2. There should be a more detailed
comparison to the earlier work of Chaudhuri et. al. In particular, they
provide results on the parameters of the DC-SBM for which there are no
errors made by their clustering algorithm and complement this with lower
bounds. At least for the case of the SBM, rather than assuming (p-r) is a
constant, a proper comparison to known results from McSherry/Chaudhuri et.
al. should be made.
3. The role of normalization (of the
embedding) is still not clear to me. While the figure shows a case where
it seems to help in practice, it would help to clarify its role in the
theory. Are the rates worse? If possible, the authors should compare the
rates with and without the normalization, and if not they should clarify
the reason for normalization.
4. Finally, in the experiments
section - you report 50% improvement after dropping 10% of the nodes.
However, although the algorithm you propose clusters *all* nodes you only
report accuracy on the retained nodes. Does the accuracy on the full set
improve at all?
Minor typos:
1. Two statistical model
--> models 2. Line 59 -- we study **an** 3. 102 -- emphasis
--> emphasize 4. 243 and 244 -- centroid -->
centroids Q2: Please summarize your review in 1-2
sentences
The paper provides an interesting theoretical analysis
for a (modification of a) widely used clustering algorithm. I have a few
issues with the theoretical results presented but I recommend
acceptance. Submitted by
Assigned_Reviewer_5
Q1: Comments to author(s).
First provide a summary of the paper, and then address the following
criteria: Quality, clarity, originality and significance. (For detailed
reviewing guidelines, see
http://nips.cc/PaperInformation/ReviewerInstructions)
NA Q2: Please summarize your review
in 1-2 sentences
NA
Submitted by
Assigned_Reviewer_6
Q1: Comments to author(s).
First provide a summary of the paper, and then address the following
criteria: Quality, clarity, originality and significance. (For detailed
reviewing guidelines, see
http://nips.cc/PaperInformation/ReviewerInstructions)
Summary: This paper analyses spectral clustering
schemes where a small positive constant (\tau) is added to the node
degrees as a form of regularization. The main contributions are the
removal of a minimum degree assumption common in previous studies and some
insights on selecting a range of good values for \tau.
Quality:
+ The theoretical part of the paper seems OK and convincing in most
parts. + The few experiments performed are evaluated in terms of
appropriate unsupervised measures such as the misclustering rate. -
The experimental section is rather weak with only one real-life dataset.
- The contributions does not seem very significant and it is unclear
why someone would prefer the proposed scheme to similar ones. - No
out-of-sample extensions are provided. - It is claimed that the paper
provides insight on the "star-shaped" figure that appears in the empirical
eigenvectors. However, these insights are not explicitly mentioned in the
manuscript. The "star" shape is known to appear in ideal conditions for
the top eigenvectors of D^{-1/2}AD^{-1/2} due to the normalization while
the top eigenvectors of D^{1}A are known to appear localized (piecewise
constant). What are the insights provided by the proposed scheme ?
Clarity: + On average, the paper is OK to read with the
following exceptions: - How are the \theta parameters computed ? -
Do the results hold for weighted graphs? Whether yes or no, it needs to be
clearly explained in the manuscript.
Originality: + The
theoretical analysis about selecting good values for \tau seems novel.
- The paper relies heavily on previous work and the contributions are
very incremental and unlikely to be significant.
Significance:
- The main contributions are rather weak. It is not clear how the lack
of the minimum degree assumption common in previous studies is significant
here. The influence of the proposed value for \tau on the clustering
results is not properly studied either. Q2: Please
summarize your review in 1-2 sentences
OK paper showing theoretical analysis on the benefits
of regularization in spectral clustering. The main contributions are on
the weak side and unlikely to be significant.
Q1:Author
rebuttal: Please respond to any concerns raised in the reviews. There are
no constraints on how you want to argue your case, except for the fact
that your text should be limited to a maximum of 6000 characters. Note
however that reviewers and area chairs are very busy and may not read long
vague rebuttals. It is in your own interest to be concise and to the
point.
Thank you to the reviewers for their thoughtful
comments.
We emphasize the following summary from Reviewer_5:
"The main contribution of the paper is to start to deal with a very
awkward assumption that has appeared in related prior work and much of the
related theoretical work on graphs with heavy-tailed degree distributions,
namely that for analytical convenience that work assumes that the minimum
degree in the graph is quite large. . . That assumption is strongly
violated in many realistic graphs, rendering prior work less than
applicable. When that assumption is removed, one gets localization on the
eigenvectors."
This paper provides the first understanding of
regularized eigenvectors for low degree nodes. Importantly, we found
(a) that the regularization is most helpful in regimes that the
previous analysis failed to address (i.e. low degree nodes). (b) the
leverage scores provide a preliminary measure of statistical uncertainty.
Previous estimation results have not provided insight at this level of
granularity; instead, previous results have studied the global estimation
performance.
Item specific responses are below.
Reviewer_4
--"need a more detailed comparison with the work of Chaudhuri et al."
Our results compare very favorably in simulation because their
estimator is often not computable. In the theory, the results are mixed.
We will rework the introduction to clarify the following points:
(Simulations) In step 4 of their algorithm, Chaudhuri et al. define a
threshold \lambda_u for each node u. In simulation, these \lambda's are
often less than 0, which makes step 5 not computable. To guarantee that
\lambda_u >0, we must have degree(u) > 64 ln(6*n / \delta). Here
delta is the probability threshold in their main theorem.
(Theory)
Our theoretical results have the advantage of not assuming anything about
the minimum degree. As Reviewer_5 observes, this is of fundamental
importance. The disadvantage of our theoretical results is that we do not
prove perfect clustering. This technical difficulty appears in all
spectral clustering results that study a canonical version of the spectral
clustering (eigendecomposition + k-means). Chaudhuri et al. circumvent
this difficulty by using half the nodes to compute singular vectors and
cluster the other half of the nodes.
Reviewer_4 --"why normalize
the rows of eigenvectors?" In a highly cited paper, Ng, Jordan and
Weiss (NIPS 2002) proposed normalizing the rows of eigenvectors and their
motivation comes from matrix perturbation theory. We provide the first
statistical analysis of this step and in fact the analysis nicely pairs
with the Degree Corrected Stochastic Blockmodel. Our revised simulation
section (discussed below) demonstrates that normalizing aids estimation
when the networks have degree heterogeneity.
Reviewer_4
--"Finally, in the experiments section - you report 50% improvement after
dropping 10% of the nodes. However, although the algorithm you propose
clusters *all* nodes you only report accuracy on the retained nodes. Does
the accuracy on the full set improve at all?" Compared with SC, it
does. The data analysis of the blog network illustrates two points:
(1) RSC is significantly better than SC on the whole dataset. (2)
The performance of RSC is more reliable on nodes with high leverage
scores, demonstrating how leverage scores provide a rough measure of
statistical uncertainty at a node-level-resolution.
Reviewer_4
--"Is the definition of mis-clustered nodes appropriate?" Because
cluster labels are not identifiable, there is no canonical measure of
"misclustered". We have adopted the definition from Rohe, Chatterjee, and
Yu (AoS 2011) because it is suitable for k-means. We will include it in
the revised supplementary material.
Reviewer_4 --"typos" Thank
you for finding and reporting the typos. We have corrected them in the
revised paper.
Reviewer_5 --"move the remarks on lines … to more
prominent position." In the revised paper, we will reorganize the
material to place greater emphasize on the following points and how they
relate to each other: a) Remove the assumption on lowest expected
degree b) Relation between leverage score and clustering performance
c) Graph sparsity and localization on eigenvectors
Reviewer_5
and 6 -- Empirical evaluation is limited. We have revised experiment 2
to more directly investigate the role of the degree distribution. We
simulate graphs with power law degree distributions with shape parameter
ranging in {4, 3.5, 3, 2.5, 2} and fixing the expected average degree to
be 8. These simulations demonstrate (1) the instability of standard
SC, (2) how normalization helps when shape parameter is less than 3,
and (3) how uniformly across simulations it is easier to cluster the
nodes with high leverage scores. These findings are consistent with
our theoretical results on introducing regularization term and
normalizing.
Reviewer_6 --"What are the insights on the
''star-shaped" pattern?" We give an model based explanation of the
"star shaped" pattern. We show that under the DC-SBM, or the extended
planted partition model, the population top eigenvectors shows exact "star
shape" (see figure 1 left panel.). And our theoretical result shows that
the empirical top eigenvectors are close to these population ones. We
will add detailed explanation in the DC-SBM section in the revised paper.
Reviewer_6 --"How are the \theta parameters computed?" In
simulation, we generate the \theta parameters from power-law distribution.
We will clarify this in the paper.
Reviewer_6 --"Do the results
hold for weighted graphs?" This is an interesting and relevant
question. We suspect that it does, but due to space constraints we cannot
address the issue here.
| |