DeepWalk: Online Learning of Social Representations

• p. 4 Section 4.1

  it is not necessary for the algorithm to work as we will show in 4.2.2

• p. 5 Section 4.4.1 Streaming

  If the cardinality of V is known (or can be bounded), we can build the Hierarchical Softmax tree for that maximum value.

• p. 4 Section 4.1 Overview

  it is not necessary for the algorithm to work as we will show in 4.2.2.

Only the authors can say what is right here.

Judge's note: Read Section 4.2.2 on the page 4 image: it describes the binary tree over V and says Huffman coding by frequency can speed up training. That implies frequencies are optional but does not address an unknown V. Section 4.4.1 on page 5 covers building the tree for a bounded |V| and assigning vertices as they are first seen. The pointer is at best partial; the authors may have meant 4.4.1.

Second check: another model agreed

• p. 1 Abstract

  F1 scores up to 10% higher than competing methods when labeled data is sparse

• p. 2 Section 1, contributions

  getting improvements 5%-10% of Micro F1, on the sparsest problems we consider

• p. 6 Section 6.1.3

  the Micro-F1 improves by 14%

• p. 6 Section 6.1.3 YouTube

  When 1% of the labeled nodes are used for test, the Micro-F1 improves by 14%

Only the authors can say what is right here.

Judge's note: This cluster is judged on the 10% versus 14% point only; the "used for test" wording that some reviewers also raised here is judged in C013. Checked pages 1, 2, 6 and Table 4 on page 7. The 14% in the text matches Table 4 against EdgeCluster (14.05 points). Against the best baseline in that column (wvRN, 26.79) the gap is 11.16, still above 10. The Macro-F1 gap is 9.74, but the introduction names Micro F1. The abstract and introduction do not say which baseline they compare against, and they understate the gain, so this is a consistency question for the authors, not a wrong result.

Second check: another model agreed

• p. 6 Section 6.1.2

  only 3% of the graph is labeled beats all other methods even when they have been given 10% of the data. In other words, DeepWalk can outperform the baselines with 60% less training data.

• p. 1 Abstract

  outperform all baseline methods while using 60% less training data

• p. 2 Section 1, contributions

  even when given 60% less training data

• p. 6 Section 6.1.2 Flickr

  only 3% of the graph is labeled beats all other methods even when they have been given 10% of the data. In other words, DeepWalk can outperform the baselines with 60% less training data

Only the authors can say what is right here.

Judge's note: Checked page 6 and Table 3 on page 7. DeepWalk Micro-F1 at 3% is 35.9, above the best baseline at 10% (SpectralClustering, 35.41); at 2% it is 34.6, below. So the 3% versus 10% comparison holds and is a 70% reduction. "60% less" would be 4% versus 10%, which Table 3 also supports (36.7 > 35.41), so the 60% claim is true as a conservative statement but does not restate the sentence it follows with "In other words". Only the authors can say which figure was meant.

Second check: another model agreed

• p. 3 Section 3.3

  introduce a mapping function Φ: v ∈ V ↦ R|V |×d

• p. 4 Section 4.2.1

  vector Φ(vj) ∈ Rd

• p. 3 3.3

  introduce a mapping function

• p. 4 4.2.1

  representation vector

Judge's note: Read on the page 3 image: the mapping is typeset with codomain R^{|V|×d} for an element v ∈ V. The next sentence says Φ is represented by a |V| × d matrix of free parameters, which is the whole parameter matrix, and page 4 writes Φ(v_j) ∈ R^d. The codomain for one vertex should be R^d. Loose notation; the meaning is recoverable.

Second check: another model agreed

• p. 3 Section 3.1

  is a vertex chosen at random from the neigh-

Judge's note: Read on the page 3 image. The walk's variables are written W^1_{v_i}, ..., W^k_{v_i}, and then the next step is drawn from the neighbours of "vertex v_k", a symbol not introduced. It is presumably shorthand for the vertex visited at step k; the intent is clear, the notation is loose.

Second check: another model agreed

• p. 2 Section 1, last paragraph

  We outline ours experiments in Section

• p. 2 Introduction, last paragraph

  We outline ours experiments in Section 5

Judge's note: Read on the page 2 image, last paragraph of Section 1. The text is as quoted; a plain grammatical slip.

Second check: none (judge only)

• p. 3 Section 3.3, first display

  W n 1 = (w0, w1, · · · , wn)

• p. 3 Section 3.3

  = (w0, w1, · · · , wn)

Judge's note: Read on the page 3 image: the subscript is 1 and the first listed element is w_0, and the following line uses Pr(w_n | w_0, w_1, ..., w_{n-1}). An index inconsistency with no effect on meaning.

Second check: none (judge only)

• p. 3 Figure 2 caption

  The power-law distribution of vertices appearing in short random walks (2a) follows a power-law

Judge's note: Read on the page 3 image; the caption is as quoted. Wording only.

Second check: none (judge only)

• p. 3 Section 3.3, first sentence

  The goal of language modeling is estimate the likelihood

• p. 3 Section 3.3, fourth paragraph

  These walks can be thought of short sentences and phrases

Judge's note: Both sentences read on the page 3 image, right column, as quoted. Two missing words.

Second check: none (judge only)

• p. 4 Algorithm 1, line 3

  for i = 0 to γ do

• p. 4 Section 4.2

  The outer loop specifies the number of times, γ, which we should start random walks at each vertex.

Judge's note: Read on the page 4 image. Taken inclusively the loop runs γ+1 times; a half-open reading (0 to γ-1) would give γ. Pseudocode off-by-one, either reading possible.

Second check: none (judge only)

• p. 6 Section 5.2, EdgeCluster bullet

  Its has been shown to perform comparably

• p. 6 Section 5.2, EdgeCluster bullet

  Its has

Judge's note: Visible on the page 6 image: "Its has been shown". Compare the wvRN bullet below, which prints "It has shown". Plain typo.

Second check: none (judge only)

• p. 6 Table 1, Labels row

  Interests

• p. 6 5.1 Datasets, BlogCatalog and Flickr bullets

  The labels represent the topic categories provided by the authors

Judge's note: Both read on the page 6 image. The wording differs between table and text ("Interests" against "topic categories"; "interest" appears in the text for Flickr, not BlogCatalog), but the terms are loose synonyms and neither is evidently wrong. A minor consistency point only.

Second check: none (judge only)

• p. 8 Section 6.2.1, third paragraph

  The first is that there is most of the benefit is accomplished by starting γ = 30 walks per node

• p. 8 6.2.1

  starting γ = 30 walks per node

• p. 8 6.2.2

  quickly slows (γ > 10)

• p. 8 Section 6.2.1

  is that there is most of the

Judge's note: Read on the page 8 image. The garbled sentence is a plain copy-edit slip. The second point raised (6.2.1 puts most of the benefit at γ = 30, while 6.2.2 says the effect "quickly slows (γ > 10)") is not confirmed as a defect: the two statements are compatible (gains slow after 10 and are mostly complete by 30), and the plotted γ values are 1, 3, 10, 30, 50, 90.

Second check: none (judge only)

• p. 8 Section 6.2, first sentence

  of DeepWalk effect its performance on classification tasks

• p. 8 Section 6.2, first sentence

  two multi-label classifications tasks

• p. 8 Section 6.2

  effect its performance on classification tasks

Judge's note: Read on the page 8 image; the verb is printed "effect". The same sentence's "classifications tasks" is judged separately as C025.

Second check: none (judge only)

• p. 8 Section 6.2

  two multi-label classifications tasks

Judge's note: Read on the page 8 image: "we conducted experiments on two multi-label classifications tasks (Flickr, and BlogCatalog)". A distinct slip from the "effect"/"affect" one in C024, though in the same sentence.

Second check: none (judge only)

• p. 8 Section 6.2.1

  dimensionality and training rate.

Judge's note: Page 8 image for the sentence; pages 6 (Sections 6.1.1-6.1.3) use "training ratio (T_R)" throughout. The meaning is clear from the context and the figure legends ("Training" 0.01-0.9), so this is a terminology inconsistency only.

Second check: none (judge only)

• p. 9 Reference [14]

  Transactions on, (6):721–741, 1984.

• p. 10 Reference [18]

  (3):267–287, 1983.

Judge's note: Read on the page 9 and page 10 images. Other journal entries in the list give volume and issue, e.g. [7] "20(1):30–42" and [11] "19(3):355–369". The omission is visible in the PDF itself; the missing volume numbers are not in the files and are not asserted here.

Second check: none (judge only)

• p. 9 Reference [3]

  Y. Bengio, R. Ducharme, and P. Vincent. A neural probabilistic language model.

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: that the 2003 JMLR paper "A neural probabilistic language model" has four authors (Bengio, Ducharme, Vincent, Jauvin). The page 9 image shows the entry as quoted, with three authors and the JMLR volume and pages. The author list of the cited work cannot be checked from the files, so this is a query for the authors.

Second check: none (judge only)

• p. 10 Reference [37]

  P. Sen, G. Namata, M. Bilgic, L. Getoor, B. Galligher,

• p. 9 Reference [12]

  B. Gallagher and T. Eliassi-Rad

• p. 10 Reference [37]

  B. Galligher

• p. 10

  B. Galligher, and T. Eliassi-Rad.

Judge's note: Page 10 image shows "B. Galligher" in [37] and "B. Gallagher" in [16]; page 9 image shows "B. Gallagher" in [12] and [13], each time alongside T. Eliassi-Rad. The inconsistency is internal to the reference list, so the fix does not depend on outside knowledge.

Second check: none (judge only)

• p. 10 Reference [44]

  An information flow modelfor conflict and fission in small groups1.

• p. 10 Reference [44]

  An information flow modelfor conflict

Judge's note: Viewed the crop of PDF page 10. Both "modelfor" and "groups1" are printed in the rendered page, so they are not extraction artefacts. The "1" is set on the baseline as part of the title; that it derives from a footnote mark in the original title is a plausible guess but not checkable here. Cosmetic.

Second check: none (judge only)