This paper won the 2018 SEOUL TEST OF TIME AWARD: http://www.iw3c2.org/updates/CP_TheWebConferenceSeoul-ToT-Award-VENG-2018.pdf ### Concerns about "primary pair": Does a primary pair exist for each n-ary relation? The following example of "NobelPrize" and "AlbertEinstein" given in the paper does not work for "Marie Curie" who won two Nobel prizes according to Wikipedia: https://en.wikipedia.org/wiki/Marie_Curie " However, this method cannot deal with additional arguments to relations that were designed to be binary. The YAGO model offers a simple solution to this problem: It is based on the assumption that for each n-ary relation, a primary pair of its arguments can be identified. For example, for the above won-prize-in-year-relation, the pair of the person and the prize could be considered a primary pair. The primary pair can be represented as a binary fact with a fact identifier: #1 : AlbertEinstein hasWonPrize NobelPrize All other arguments can be represented as relations that hold between the primary pair and the other argument: #2 : #1 time 1921 "

### Concerns about the time complexity given in Table 1: Section 3.3.5. Graph Factorization (GF). "To the best of our knowledge, Graph Factorization [21] was the first method to obtain a graph embedding in $O(|E|)$ time". This contradicts the time complexities given in Table 1 "$O(|E|d)$" for GF and earlier complexities in $O(|E|d^2)$. ### Dimensionality reduction is different from graph embedding: Some of the given references are actually for dimensionality reductions: the input is a set of vectors in dimension $D$ and the output is a set of vectors in dimension $d$ (such that $d<<D$). The problem where the input is a graph and the output is a set of vectors of small dimension is actually different. LLE and Laplacian Eigenmaps are methods starting from a set of vectors, then forming a graph out of the vectors and then forming vectors of small dimension out of the graph. Maybe this can be emphasized. In particular, maybe using these techniques on social networks (for instance) may not be adapted. The input graph is in most of the cases unweighted and the structure might be very different from the structure of a graph obtained from vectors. ### Cauchy graph embedding: - What is the complexity of Cauchy graph embedding (not given in table 1)? - How is the solution obtained? From eigenvectors like the two previous methods (LLE and Laplacian Eigenmaps)? - How to tune the parameter $\sigma$? After reading the paper [33], the Cauchy graph embedding method is more complex and seems less scalable. ### GF and SDP matrices: Page 4, section 3.3.5. Graph Factorization (GF). "Note that as the adjacency matrix is often not positive semidefinite, the minimum of the loss function is greater than 0 even if the dimensionality of embedding is |V|.". - The loss function is never 0 if $\lambda>0$. - If $W$ is SDP (and different from the null matrix) then the associated graph needs to contain self-loops. If self-loops are forbidden, then the matrix is never SDP. - If $W$ is SDP, then there exist a set of vectors $\{Y_i\}$ such that $\forall i,j$, $W_{ij}=<Y_i,Y_j>$. Is it true that such vectors can always be of dimension $n$ (the size of $W$) or less? A reference or proof? ### Minors: - Page 3, section 3.1.1. Locally Linear Embedding (LLE). It should be specified that the rows of the weight matrix need sum to one: $\sum_j W_{ij}=1$. - Page 3, section 3.1.1. Locally Linear Embedding (LLE). "the sparse matrix $(I-W)^T(I-W)$" this matrix is often not sparse (in the sense that it has a lot more non-zero values compared to the adjacency matrix $W$). - Page 3, section 3.1.3. Cauchy Graph Embedding. Is there a solution of the optimization using eigenvectors as in the two previous methods? - What is "positive pointwise mutual information (PPMI)"? - Figure 1 refers to CPE (Community Preserving Embedding), but CPE is not discussed in the survey. - The section on GraRep refers to the section on HOPE which is after, maybe swapping the two sections? ### Typos: - Many text-overflows