NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:2699
Title:Hyperbolic Graph Convolutional Neural Networks

Reviewer 1

The paper is well written in general although it contains mistakes and ignores some related work. In particular, it is not clear whether the corollaries (whose proofs are given in the appendix) are sold as contributions or not. Many of their implications are already known in the machine learning literature (see details below). == Mistakes: - Wrong curvature (minor mistake): The hyperboloid defined in Eq. (3) is said to have a constant curvature of -1/K^2 in the submission. As explained in detail in Section 3.4 of Chapter 3 of the second edition of [1A] (or also in the following references [1C] and [1D]), its curvature is actually -1/K. The explanation in these references uses the fact that the hyperboloid can be seen as a half spherical model with imaginary radius equal to i * sqrt(K). From this duality between the spherical and hyperbolic models, many formulations (e.g. exponential and logarithm maps, parallel transport etc...) can be infered from one model to the other (see details for the mistake below). - Wrong parallel transport formulation (major mistake): The parallel transport formula in Eq. (15) (supp. material) is wrong. The denominator of the right term should be the squared geodesic distance instead of the geodesic distance. The details can for example be found in the last equation of Section B.1 of [1B], which is given for the spherical case although the hyperbolic case is very similar due to the duality between the spherical and hyperbolic models. This can be verified by observing that the Minkowski inner product between examples belonging to tangent space (e.g. of the north pole in H^{d,1} in the paper) should be preserved via parallel transport. I provide a code in Matlab below to show that the provided parallel transport equation is wrong. If the parallel transport formula in Eq. (15) was correct, then r1 would be equal to r3 (obtained by using Eq. (15)). However, r1 and r2 (obtained by using the hyperbolic adaptation of [1B]) are equal. I did not manage to check if the source code also exploited that formulation of the parallel transport. == Connections to existing works It is not clear if the different theorems in Section 3 (Background section) are sold as contributions since the proofs are given. These results either correspond to definitions or are already known in the literature. For instance, Corollary 3.1 contains definitions and seems to be useful to obtain a Poincare distance that depends on the curvature. The version of the Poincare distance that depends on the curvature (Eq. (4) of the submission) is already known in the machine learning literature. It is for example given in Eq. (18) of [1C]. The hyperbolic exponential and logarithm maps that depend on the curvature are also given in Section 6.1 of [1C]. Theorem 4.1 is very similar to the last paragraph of the proof given in Section A.2 of [1D]. Law et. al [1D] also observe that decreasing the curvature of the hyperbolic space improves classification and retrieval performances. Nonetheless, they do not exploit multiple layers of hyperbolic representations. == Experiments and baselines The experimental results (node classification (with F1 score) or unseen link predictions between nodes (with ROC evaluation metric)) show huge improvement of the hyperbolic GCN when each layer has a specific curvature. The comparison to Euclidean baselines is not really fair since their architecture contains many more hyperparameters which are: - the hyperboloid radii (one for each layer of the GCN) - the bias r in Eq. (14) - the temperature t in Eq. (14) Each curvature can be seen as a hyperparameter, this curvature is equal to 0 in the Euclidean case. The paper mentions in line 225 that having different curvatures is very important for good performance. Fortunately, Table 2 shows an ablation study with less tuning of the hyperparameters. In conclusion, the main contributions of the submission then seem to be only the adaptation of GCNs to the hyperbolic case, and huge experimental improvement compared to Euclidean baselines. minor comments: - the MLP acronym is first used in line 186 although it is defined in line 272. - there is already a model named HyperGCN (Yadati et al.) [1A] Peter Petersen, Riemannian Geometry, 2006 [1B] Bergmann, R., Fitschen, J. H., Persch, J., & Steidl, G. , Priors with coupled first and second order differences for manifold-valued image processing, Journal of mathematical imaging and vision, 60(9) [1C] Wilson, R., Hancock, E., Pekalska, E., Duin, R., Spherical and Hyperbolic Embeddings of Data, IEEE T-PAMI, 2014 [1D] Law, M., Liao, R., Snell, J., Zemel, R., Lorentzian Distance Learning for Hyperbolic Representations, ICML 2019 (rejected by ICLR 2019, so publicly available since September 2018) [1E] Ganea O., Becigneul G., Hofmann T., Hyperbolic Neural Networks, NIPS 2018 == code showing that Eq. (15) is not correct (using a curvature of -1 for simplicity) d = 4; o = 10*rand(d-1,1); o = [sqrt(sum(o.^2) + 1);o]; %o = [1;0;0;0]; y = 10*rand(d-1,1); y = [sqrt(sum(y.^2) + 1);y]; a = create_tangent_vector(o); b = create_tangent_vector(o); r1 = minkowski_inner_product(a,b) r2 = minkowski_inner_product(parallel_transport(a, o, y), parallel_transport(b, o, y)) r3 = minkowski_inner_product(parallel_transport_submission(a, o, y), parallel_transport_submission(b, o, y)) function a = create_tangent_vector(o) a = 10*rand(length(o)-1,1); a0 = (a'*o(2:end)) / o(1); a = [a0;a]; end function r = minkowski_inner_product(x,y) e = x .* y; e(1) = -e(1); r = sum(e); end function r = hyperboloid_normalization(x) abs_norm = sqrt(minkowski_inner_product(x,x)); r = x / abs_norm; end function r = geodesic_distance(x,y) r = acosh(-minkowski_inner_product(x,y)); end function r = logarithm_map(x,y) r = y + minkowski_inner_product(y,x) * x; r = hyperboloid_normalization(r) * geodesic_distance(x,y); end function r = parallel_transport(xi, x, y, used_power) if nargin < 4 used_power = 2; end r = xi - minkowski_inner_product(logarithm_map(x,y), xi) / (geodesic_distance(x,y)^used_power) * (logarithm_map(x,y) + logarithm_map(y,x)); end function r = parallel_transport_submission(xi, x, y) r = parallel_transport(xi, x, y, 1); end edit after rebutall: The authors confirmed that the parallel transport formula they used was correct, so I am okay with it. The theoretical novelty is limited but in terms of experimental contributions, the paper is interesting.

Reviewer 2

General Comments: The paper is very well written although very dense. I enjoyed the supplementary material, it helped me cover some background knowledge that I needed to understand the paper. The material is very original and very useful since it improves the original goal of hyperbolic neural networks (model graph and tree-like data). The best way is to actually combine it with Graph NNs that we already know perform really well. The experiments are extensive and convincing line 27:” due to the fact that the number of elements/nodes increases exponentially with hierarchy, but the area in Euclidean space grows linearly with distance.” Which area exactly? The area of a disk grows quadratically with respect to the radious not linearly. Also, this argument is fundamentaly wrong as it compares an infinite space with a finite space. The number of the leafs of a hierarchical space grows exponentially but remains finite. The area of a disk contains infinite points. Technically speaking the unit circle can fit the universe. The reason why euclidean space is different and it is described very well on this paper: As described in this paper The maximum number of points which can share a common nearest neighbor is limited 2 for 1-dimensional spaces, 5 for 2-dimensional spaces while such centralized structures do exist in real data d-dimensional spaces can at most embed (d + 1) points with uniform pair-wise similarities. See the references: K. Zeger and A. Gersho. How many points in Euclidean space can have a common nearest neighbor? In International Symposium on Information Theory, page 109, 1994. L. van der Maaten and G. E. Hinton. Visualizing non-metric similarities in multiple maps. Machine Learning, 87(1):33–55, 2012. I suggest reading this blog and also this paper: line 91: “In contrast to previous work, we derive core neural network operations in a more stable model” That statement is not backed. It is a bit in the air. You have to be more specific. One of the weaknesses of this paper is that it doesn’t stress the differences with Reference [10] Hyperbolic Neural Networks. Especially on the derived operators. From my understanding, the contribution is on aggregation layers and attention mechanisms. It also seems to me that another difference is that you work on the Lorentz manifold (which you call hyperboloid, it would be better to use Lorentz, since it is more widely known as that in the ML community).

Reviewer 3

A novel hyperbolic graph convolutional network is proposed in this submission, which shows outstanding performance on the graphs possessing hierarchical structure in the link prediction and link prediction tasks. The only concern the reviewer has is the definition of activation function defined in equation 10. It means that after implementing a normal activation function on T_o^{k_{l-1}}, the intermediate vector defined in this tangent space is directly mapped to a space with radius k_l via exp_o^{k_l}, is this mapping even mathematically plausible?