NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:1222
Title:Improving Textual Network Learning with Variational Homophilic Embeddings

Reviewer 1

CLARITY: Overall the paper is clearly written. There were few parts that I had hard time understanding (i.e., such as line 223-224, how can q(z_i|x_i, x_j, w_ij) denotes the posterior with and without edge information? I guess it's when q_0, and when q_1?) The authors seemed to have squeezed the format a bit too much. There's a lot of details, might be better to focus on key contributions and make the paper self-containing. For example, Figure 2, as is, impossible to decipher and we have to look up the supplementary material to understand. Overall, I found it hard to tell apart the contribution of homophilic prior on the performance. The particular design choice of encoder and decoder architecture (section 3.2), as well as how to set the global network embeddings (section 3.3). A proper ablation study should be designed to test where the improvement is coming from. I assume maybe the PWA baseline that the author is proposing is the proposed method minus the homophilic prior? But it wasn't super clear. And then the delta between PWA and VHE is fairly small... The model introduces the notion of link, no link, and unknown, but then the experiment doesn't show whether modeling of unknown brings meaningful differences? ==== AFTER THE AUTHOR RESPONSE: Thanks for careful response and pointing out the ablations. It is much clearer with the author response. Though, about (d), the figure doesn't seem to support that modeling unknown link brings significant gain..

Reviewer 2

[I read the author response; my score remains unchanged] The writing of the paper is very clear, and the paper does a good job of motivating the method and describing the technical details of the approach. I did not find any issues with the technical derivations for the VAE formulation, and to my knowledge the proposed VHE method is novel. The proposed method seems very practical, especially because it can tolerate incomplete graphs, missing edges, and new nodes that appear at test time. The authors include their phrase-to-word alignment system among their list of contributions, but that portion of the work appears less novel. There have been a number of methods that rely on mutual attention between two sequences of tokens (see e.g. ;, and to the extent that the authors' implementation has incremental differences over the past work there is no indication that these differences affect accuracy. That said, the word representation method is not the key focus of the present paper. [Regarding the response: my initial reaction reading the paper for the first time was that the authors were claiming novelty of the architecture itself; I agree that the application of such an architecture to this specific task is separate and can be considered a contribution of this paper.] [Thank you for promising to address some of the presentation issues below] 77: Why is it "without loss of generality"? I don't follow. 190: Are K_r/K_c a function of L_j/L_i, or are they fixed hyperparameters? I think "number of filters" generally refers to the channel dimension, but these are sequence length dimensions instead. In general the text in 183-195 isn't as clear as the remainder of the paper. Grammar and formatting: 20: its -> their 265: What does PWA stand for? Table 2: Could you say more about the methods? For example, adding columns to indicate whether each method encodes topology, content, both, is a discriminative method, is a generative method. Defining the acronyms might also help. 295: "When VHE works?" -> "When does VHE work?" 344: grammar seems a bit odd at the phrase "are not necessary". Maybe "documents with similar info do not necessarily cite each other", or "documents citing each other do not necessarily have similar info" (depending on what the intended meaning is) [349: textural -> textual (?)]

Reviewer 3

This paper uses the formalism of autoencoders to model the joint probability distribution of the text attributes of a _ pair of nodes _. The correlation/dependence is encapsulated by the corresponding pair of latent variables, in the case where the two nodes are connected by an edge. In a first derivation (7), the edge is assumed visible. Later (8), it is assumed latent and marginalized. The manifest advantage of this approach is that the information about edges and the text attributes of the vertices are encoded simultaneosuly, unlike previous approaches as reviewed by the authors. The parameters of the posterior distributions are estimated using an original textual embedding which takes into account the text of the other nodes (Section 3.2). Even just this embedding alone leads to significant accuracy improvements for an existing SOTA method (CANE; the integrated method is noted as PWA in the tables). I only have a somehow major remark about the presented approach. The authors claim that this method is able to encode "structural information" per node. This means, create an embedding of its connectivity to other nodes. In reality, the embedding (h) used in the approach is just an arbitrary vector per node that is learned by minimizing the training objective. How could this be proved to be connectivity information? The authors also state that they are able to mollify the computational complexity deriving from modelling pairs of vertices. It would be very useful if they added actual computational times for both training and inference (they are missing also from the Supplementary). ==== AFTER THE AUTHOR RESPONSE: I am overall satisfied with the authors' response. About H, I had no doubt that it is useful (see Fig. 4(d)): whether it comes into correspondence with actual structural information is still to be proved.