Published as a conference paper at ICLR 2022

H OW ATTENTIVE ARE G RAPH ATTENTION

N ETWORKS ?
Shaked Brody Uri Alon
Technion Language Technologies Institute
[email protected] Carnegie Mellon University
[email protected]

Eran Yahav
Technion
arXiv:2105.14491v3 [cs.LG] 31 Jan 2022

[email protected]

A BSTRACT

Graph Attention Networks (GATs) are one of the most popular GNN architectures
and are considered as the state-of-the-art architecture for representation learning
with graphs. In GAT, every node attends to its neighbors given its own represen-
tation as the query. However, in this paper we show that GAT computes a very
limited kind of attention: the ranking of the attention scores is unconditioned
on the query node. We formally define this restricted kind of attention as static
attention and distinguish it from a strictly more expressive dynamic attention. Be-
cause GATs use a static attention mechanism, there are simple graph problems
that GAT cannot express: in a controlled problem, we show that static attention
hinders GAT from even fitting the training data. To remove this limitation, we
introduce a simple fix by modifying the order of operations and propose GATv2:
a dynamic graph attention variant that is strictly more expressive than GAT. We
perform an extensive evaluation and show that GATv2 outperforms GAT across 12
OGB and other benchmarks while we match their parametric costs. Our code is
available at https://github.com/tech-srl/how_attentive_are_
gats.1 GATv2 is available as part of the PyTorch Geometric library,2 the Deep
Graph Library,3 and the TensorFlow GNN library.4

1 I NTRODUCTION

Graph neural networks (GNNs; Gori et al., 2005; Scarselli et al., 2008) have seen increasing popularity
over the past few years (Duvenaud et al., 2015; Atwood and Towsley, 2016; Bronstein et al., 2017;
Monti et al., 2017). GNNs provide a general and efficient framework to learn from graph-structured
data. Thus, GNNs are easily applicable in domains where the data can be represented as a set of
nodes and the prediction depends on the relationships (edges) between the nodes. Such domains
include molecules, social networks, product recommendation, computer programs and more.
In a GNN, each node iteratively updates its state by interacting with its neighbors. GNN variants
(Wu et al., 2019; Xu et al., 2019; Li et al., 2016) mostly differ in how each node aggregates and
combines the representations of its neighbors with its own. Veličković et al. (2018) pioneered the use
of attention-based neighborhood aggregation, in one of the most common GNN variants – Graph
Attention Network (GAT). In GAT, every node updates its representation by attending to its neighbors
using its own representation as the query. This generalizes the standard averaging or max-pooling
of neighbors (Kipf and Welling, 2017; Hamilton et al., 2017), by allowing every node to compute
a weighted average of its neighbors, and (softly) select its most relevant neighbors. The work of
1
An annotated implementation of GATv2 is available at https://nn.labml.ai/graphs/gatv2/
2
from torch_geometric.nn.conv.gatv2_conv import GATv2Conv
3
from dgl.nn.pytorch import GATv2Conv
4
from tensorflow_gnn.graph.keras.layers.gat_v2 import GATv2Convolution

1
Published as a conference paper at ICLR 2022

k0 k1 k2 k3 k4 k5 k6 k7 k8 k9 k0 k1 k2 k3 k4 k5 k6 k7 k8 k9
q0 0.08 0.10 0.10 0.07 0.08 0.08 0.11 0.09 0.20 0.08 q0 0.95 0.00 0.00 0.01 0.01 0.00 0.00 0.02 0.01 0.00
q1 0.05 0.10 0.10 0.04 0.04 0.04 0.13 0.06 0.38 0.04 q1 0.01 0.92 0.01 0.01 0.01 0.00 0.01 0.01 0.00 0.02
q2 0.05 0.10 0.10 0.04 0.05 0.05 0.13 0.06 0.38 0.05 q2 0.00 0.00 0.95 0.00 0.00 0.01 0.02 0.01 0.00 0.00
q3 0.08 0.10 0.10 0.07 0.08 0.08 0.10 0.09 0.24 0.08 q3 0.01 0.01 0.00 0.94 0.00 0.01 0.00 0.00 0.02 0.01
q4 0.08 0.09 0.09 0.07 0.07 0.07 0.10 0.08 0.27 0.07 q4 0.00 0.00 0.00 0.00 0.96 0.00 0.00 0.01 0.01 0.00
q5 0.09 0.11 0.11 0.08 0.09 0.08 0.11 0.10 0.16 0.09 q5 0.00 0.01 0.01 0.01 0.01 0.89 0.01 0.01 0.04 0.02
q6 0.04 0.10 0.11 0.03 0.04 0.04 0.14 0.06 0.40 0.04 q6 0.00 0.01 0.04 0.00 0.01 0.01 0.86 0.02 0.01 0.03
q7 0.07 0.09 0.09 0.06 0.07 0.07 0.10 0.08 0.29 0.07 q7 0.04 0.02 0.01 0.01 0.03 0.01 0.00 0.87 0.00 0.01
q8 0.04 0.11 0.11 0.02 0.04 0.03 0.14 0.07 0.41 0.04 q8 0.01 0.00 0.01 0.01 0.01 0.01 0.01 0.00 0.94 0.00
q9 0.07 0.09 0.09 0.06 0.07 0.07 0.11 0.08 0.30 0.07 q9 0.01 0.02 0.01 0.01 0.01 0.01 0.01 0.00 0.00 0.93

1.0
0.4 q0
q1
q2 0.8 q0
0.3 q3 q1
q4 q2
q5 0.6
q6 q3
0.2 q7 q4
q8 0.4 q5
q9 q6
q7
0.1 0.2 q8
q9
0.0
k0 k1 k2 k3 k4 k5 k6 k7 k8 k9 k0 k1 k2 k3 k4 k5 k6 k7 k8 k9
(a) Attention in standard GAT (Veličković et al. (2018)) (b) Attention in GATv2, our fixed version of GAT

Figure 1: In a complete bipartite graph of “query nodes” {q0, ..., q9} and “key nodes” {k0, ..., k9}:
standard GAT (Figure 1a) computes static attention – the ranking of attention coefficients is global
for all nodes in the graph, and is unconditioned on the query node. For example, all queries (q0 to
q9) attend mostly to the 8th key (k8). In contrast, GATv2 (Figure 1b) can actually compute dynamic
attention, where every query has a different ranking of attention coefficients of the keys.

Veličković et al. also generalizes the Transformer’s (Vaswani et al., 2017) self-attention mechanism,
from sequences to graphs (Joshi, 2020).
Nowadays, GAT is one of the most popular GNN architectures (Bronstein et al., 2021) and is
considered as the state-of-the-art neural architecture for learning with graphs (Wang et al., 2019a).
Nevertheless, in this paper we show that GAT does not actually compute the expressive, well known,
type of attention (Bahdanau et al., 2014), which we call dynamic attention. Instead, we show that
GAT computes only a restricted “static” form of attention: for any query node, the attention function
is monotonic with respect to the neighbor (key) scores. That is, the ranking (the argsort) of attention
coefficients is shared across all nodes in the graph, and is unconditioned on the query node. This fact
severely hurts the expressiveness of GAT, and is demonstrated in Figure 1a.
Supposedly, the conceptual idea of attention as the form of interaction between GNN nodes is
orthogonal to the specific choice of attention function. However, Veličković et al.’s original design of
GAT has spread to a variety of domains (Wang et al., 2019a; Yang et al., 2020; Wang et al., 2019c;
Huang and Carley, 2019; Ma et al., 2020; Kosaraju et al., 2019; Nathani et al., 2019; Wu et al., 2020;
Zhang et al., 2020) and has become the default implementation of “graph attention network” in all
popular GNN libraries such as PyTorch Geometric (Fey and Lenssen, 2019), DGL (Wang et al.,
2019b), and others (Dwivedi et al., 2020; Gordić, 2020; Brockschmidt, 2020).
To overcome the limitation we identified in GAT, we introduce a simple fix to its attention function
by only modifying the order of internal operations. The result is GATv2 – a graph attention variant

2
Published as a conference paper at ICLR 2022

that has a universal approximator attention function, and is thus strictly more expressive than GAT.
The effect of fixing the attention function in GATv2 is demonstrated in Figure 1b.
In summary, our main contribution is identifying that one of the most popular GNN types, the graph
attention network, does not compute dynamic attention, the kind of attention that it seems to compute.
We introduce formal definitions for analyzing the expressive power of graph attention mechanisms
(Definitions 3.1 and 3.2), and derive our claims theoretically (Theorem 1) from the equations of
Veličković et al. (2018). Empirically, we use a synthetic problem to show that standard GAT cannot
express problems that require dynamic attention (Section 4.1). We introduce a simple fix by switching
the order of internal operations in GAT, and propose GATv2, which does compute dynamic attention
(Theorem 2). We further conduct a thorough empirical comparison of GAT and GATv2 and find
that GATv2 outperforms GAT across 12 benchmarks of node-, link-, and graph-prediction. For
example, GATv2 outperforms extensively tuned GNNs by over 1.4% in the difficult “UnseenProj
Test” set of the VarMisuse task (Allamanis et al., 2018), without any hyperparameter tuning; and
GATv2 improves over an extensively-tuned GAT by 11.5% in 13 prediction objectives in QM9. In
node-prediction benchmarks from OGB (Hu et al., 2020), not only that GATv2 outperforms GAT
with respect to accuracy – we find that dynamic attention provided a much better robustness to noise.

2 P RELIMINARIES

A directed graph G = (V, E) contains nodes V = {1, ..., n} and edges E ⊆ V × V, where (j, i) ∈ E
denotes an edge from a node j to a node i. We assume that every node i ∈ V has an initial
(0)
representation hi ∈ Rd0 . An undirected graph can be represented with bidirectional edges.

2.1 G RAPH N EURAL N ETWORKS

A graph neural network (GNN) layer updates every node representation by aggregating its neighbors’
representations. A layer’s input is a set of node representations {hi ∈ Rd | i ∈ V} and the set of
0
edges E. A layer outputs a new set of node representations {h0i ∈ Rd | i ∈ V}, where the same
parametric function is applied to every node given its neighbors Ni = {j ∈ V | (j, i) ∈ E}:
h0i = fθ (hi , AGGREGATE ({hj | j ∈ Ni })) (1)
The design of f and AGGREGATE is what mostly distinguishes one type of GNN from the other. For
example, a common variant of GraphSAGE (Hamilton et al., 2017) performs an element-wise mean
as AGGREGATE, followed by concatenation with hi , a linear layer and a ReLU as f .

2.2 G RAPH ATTENTION N ETWORKS

GraphSAGE and many other popular GNN architectures (Xu et al., 2019; Duvenaud et al., 2015)
weigh all neighbors j ∈ Ni with equal importance (e.g., mean or max-pooling as AGGREGATE). To
address this limitation, GAT (Veličković et al., 2018) instantiates Equation (1) by computing a learned
weighted average of the representations of Ni . A scoring function e : Rd × Rd → R computes a score
for every edge (j, i), which indicates the importance of the features of the neighbor j to the node i:

e (hi , hj ) = LeakyReLU a> · [W hi kW hj ]

(2)
0 0
where a ∈ R2d , W ∈ Rd ×d are learned, and k denotes vector concatenation. These attention scores
are normalized across all neighbors j ∈ Ni using softmax, and the attention function is defined as:
exp (e (hi , hj ))
αij = softmaxj (e (hi , hj )) = P (3)
j 0 ∈N
i
exp (e (hi , hj 0 ))
Then, GAT computes a weighted average of the transformed features of the neighbor nodes (followed
by a nonlinearity σ) as the new representation of i, using the normalized attention coefficients:
X 
h0i = σ αij · W hj (4)
j∈Ni

From now on, we will refer to Equations (2) to (4) as the definition of GAT.

3
Published as a conference paper at ICLR 2022

3 T HE E XPRESSIVE P OWER OF G RAPH ATTENTION M ECHANISMS

In this section, we explain why attention is limited when it is not dynamic (Section 3.1). We then
show that GAT is severely constrained, because it can only compute static attention (Section 3.2).
Next, we show how GAT can be fixed (Section 3.3), by simply modifying the order of operations.
We refer to a neural architecture (e.g., the scoring or the attention function of GAT) as a family of
functions, parameterized by the learned parameters. An element in the family is a concrete function
with specific trained weights. In the following, we use [n] to denote the set [n] = {1, 2, ..., n} ⊂ N.

3.1 T HE I MPORTANCE OF DYNAMIC W EIGHTING

Attention is a mechanism for computing a distribution over a set of input key vectors, given an
additional query vector. If the attention function always weighs one key at least as much as any other
key, unconditioned on the query, we say that this attention function is static:
Definition 3.1
(Static attention). A (possibly infinite) family of scoring functions F ⊆
Rd × Rd → R computes static scoring for a given set of key vectors K = {k1 , ..., kn } ⊂ Rd and
query vectors Q = {q1 , ..., qm } ⊂ Rd , if for every f ∈ F there exists a “highest

scoring” key jf ∈ [n]
such that for every query i ∈ [m] and key j ∈ [n] it holds that f qi , kjf ≥ f (qi , kj ). We say
that a family of attention functions computes static attention given K and Q, if its scoring function
computes static scoring, possibly followed by monotonic normalization such as softmax.

Static attention is very limited because every function f ∈ F has a key that is always selected,
regardless of the query. Such functions cannot model situations where different keys have different
relevance to different queries. Static attention is demonstrated in Figure 1a.
The general and powerful form of attention is dynamic attention:
Definition 3.2
(Dynamic attention). A (possibly infinite) family of scoring functions F ⊆
Rd × Rd → R computes dynamic scoring for a given set of key vectors K = {k1 , ..., kn } ⊂ Rd
and query vectors Q = {q1 , ..., qm } ⊂ Rd , if for any mapping ϕ:
[m] → [n] there exists f ∈ F such
that for any query i ∈ [m] and any key j6=ϕ(i) ∈ [n]: f qi , kϕ(i) > f (qi , kj ). We say that a family
of attention functions computes dynamic attention for K and Q, if its scoring function computes
dynamic scoring, possibly followed by monotonic normalization such as softmax.


That is, dynamic attention can select every key ϕ (i) using the query i, by making f qi , kϕ(i) the
maximal in {f (qi , kj ) | j ∈ [n]}. Note that dynamic and static attention are exclusive properties,
but they are not complementary. Further, every dynamic attention family has strict subsets of static
attention families with respect to the same K and Q. Dynamic attention is demonstrated in Figure 1b.

Attending by decaying Another way to think about attention is the ability to “focus” on the most
relevant inputs, given a query. Focusing is only possible by decaying other inputs, i.e., giving these
decayed inputs lower scores than others. If one key is always given an equal or greater attention score
than other keys (as in static attention), no query can ignore this key or decay this key’s score.

3.2 T HE L IMITED E XPRESSIVITY OF GAT

Although the scoring function e can be defined in various ways, the original definition of Veličković
et al. (2018) (Equation (2)) has become the de facto practice: it has spread to a variety of domains and
is now the standard implementation of “graph attention network” in all popular GNN libraries (Fey
and Lenssen, 2019; Wang et al., 2019b; Dwivedi et al., 2020; Gordić, 2020; Brockschmidt, 2020).
The motivation of GAT is to compute a representation for every node as a weighted average of its
neighbors. Statedly, GAT is inspired by the attention mechanism of Bahdanau et al. (2014) and the
self-attention mechanism of the Transformer (Vaswani et al., 2017). Nonetheless:
Theorem 1. A GAT layer computes only static attention, for any set of node representations K =
Q = {h1 , ..., hn }. In particular, for n > 1, a GAT layer does not compute dynamic attention.

Proof. Let G = (V, E) be a graph modeled by a GAT layer with some a and W values (Equations (2)
and (3)), and having node representations {h1 , ..., hn }. The learned parameter a can be written as a

4
Published as a conference paper at ICLR 2022

0 0
concatenation a = [a1 ka2 ] ∈ R2d such that a1 , a2 ∈ Rd , and Equation (2) can be re-written as:

e (hi , hj ) = LeakyReLU a> >

1 W hi + a2 W hj (5)

Since V is finite, there exists a node jmax ∈ V such that a> 2 W hjmax is maximal among all nodes
j ∈ V (jmax is the jf required by Definition 3.1). Due to the monotonicity of LeakyReLU and
softmax, for every query node i ∈ V, the node jmax also leads to the maximal value of its attention
distribution {αij | j ∈ V}. Thus, from Definition 3.1 directly, α computes only static attention. This
also implies that α does not compute dynamic attention, because in GAT, Definition 3.2 holds only
for constant mappings ϕ that map all inputs to the same output.

The consequence of Theorem 1 is that for any set of nodes V and a trained GAT layer, the attention
function α defines a constant ranking (argsort) of the nodes, unconditioned on the query nodes
i. That is, we can denote sj = a> 2 W hj and get that for any choice of hi , α is monotonic with
respect to the per-node scores {sj | j ∈ V}. This global ranking induces the local ranking of every
neighborhood Ni . The only effect of hi is in the “sharpness” of the produced attention distribution.
This is demonstrated in Figure 1a (bottom), where different curves denote different queries (hi ).

Generalization to multi-head attention Veličković et al. (2018) found it beneficial to employ H

separate attention heads and concatenate their outputs, similarly to Transformers. In this case,
Theorem 1 holds for each head separately: every head h ∈ [H] has a (possibly different) node that
(h)
maximizes {sj | j ∈ V} , and the output is the concatenation of H static attention heads.

3.3 B UILDING DYNAMIC G RAPH ATTENTION N ETWORKS

To create a dynamic graph attention network, we modify the order of internal operations in GAT and
introduce GATv2 – a simple fix of GAT that has a strictly more expressive attention mechanism.
GATv2 The main problem in the standard GAT scoring function (Equation (2)) is that the learned
layers W and a are applied consecutively, and thus can be collapsed into a single linear layer. To fix
this limitation, we simply apply the a layer after the nonlinearity (LeakyReLU), and the W layer
after the concatenation, 5 effectively applying an MLP to compute the score for each query-key pair:
e (hi , hj ) =LeakyReLU a> · [W hi kW hj ]


GAT (Veličković et al., 2018): (6)
GATv2 (our fixed version): e (hi , hj ) =a> LeakyReLU (W · [hi khj ]) (7)
The simple modification makes a significant difference in the expressiveness of the attention function:
Theorem 2. A GATv2 layer computes dynamic attention for any set of node representations K =
Q = {h1 , ..., hn } .

We prove Theorem 2 in Appendix A. The main idea is that we can define an appropriate function
that GATv2 will be a universal approximator (Cybenko, 1989; Hornik, 1991) of. In contrast, GAT
(Equation (52)) cannot approximate any such desired function (Theorem 1).
Complexity GATv2 has the same time-complexity as GAT’s declared complexity: O (|V|dd0 + |E|d0 ).
However, by merging its linear layers, GAT can be computed faster than stated by Veličković et al.
(2018). For a detailed time- and parametric-complexity analysis, see Appendix G.

4 E VALUATION

First, we demonstrate the weakness of GAT using a simple synthetic problem that GAT cannot even fit
(cannot even achieve high training accuracy), but is easily solvable by GATv2 (Section 4.1). Second,
we show that GATv2 is much more robust to edge noise, because its dynamic attention mechanisms
allow it to decay noisy (false) edges, while GAT’s performance severely decreases as noise increases
(Section 4.2). Finally, we compare GAT and GATv2 across 12 benchmarks overall. (Sections 4.3
to 4.6 and appendix D.3). We find that GAT is inferior to GATv2 across all examined benchmarks.
5
We also add a bias vector b before applying the nonlinearity, we omit this in Equation (7) for brevity.

5
Published as a conference paper at ICLR 2022

A,? B,? C,? D,? 100 GATv2 test

90 GAT8h test
80
GAT1h train
70

Accuracy
60
GAT1h test
50
A,4 B,3 C,2 D,1 40
30
Figure 2: The D ICTIONARY- 20
L OOKUP problem of size k=4: ev- 10
0
ery node in the bottom row has an 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
alphabetic attribute ({A, B, C, ...})
and a numeric value ({1, 2, 3, ...}); k (number of different keys in each graph)
every node in the upper row has
only an attribute; the goal is to pre- Figure 3: The D ICTIONARY L OOKUP problem: GATv2 easily
dict the value for each node in the achieves 100% train and test accuracies even for k=100 and
upper row, using its attribute. using only a single head.

Setup When previous results exist, we take hyperparameters that were tuned for GAT and use them
in GATv2, without any additional tuning. Self-supervision (Kim and Oh, 2021; Rong et al., 2020a),
graph regularization (Zhao and Akoglu, 2020; Rong et al., 2020b), and other tricks (Wang, 2021;
Huang et al., 2021) are orthogonal to the contribution of the GNN layer itself, and may further improve
all GNNs. In all experiments of GATv2, we constrain the learned matrix by setting W = [W 0 kW 0 ],
to rule out the increased number of parameters over GAT as the source of empirical difference (see
Appendix G.2). Training details, statistics, and code are provided in Appendix B.
Our main goal is to compare dynamic and static graph attention mechanisms. However, for reference,
we also include non-attentive baselines such as GCN (Kipf and Welling, 2017), GIN (Xu et al., 2019)
and GraphSAGE (Hamilton et al., 2017). These non-attentive GNNs can be thought of as a special
case of attention, where every node gives all its neighbors the same attention score. Additional
comparison to a Transformer-style scaled dot-product attention (“DPGAT”), which is strictly weaker
than our proposed GATv2 (see a proof in Appendix E.1), is shown in Appendix E.

4.1 S YNTHETIC B ENCHMARK : D ICTIONARY L OOKUP

The D ICTIONARY L OOKUP problem is a contrived problem that we designed to test the ability of
a GNN architecture to perform dynamic attention. Here, we demonstrate that GAT cannot learn
this simple problem. Figure 2 shows a complete bipartite graph of 2k nodes. Each “key node” in
the bottom row has an attribute ({A, B, C, ...}) and a value ({1, 2, 3, ...}). Each “query node” in
the upper row has only an attribute ({A, B, C, ...}). The goal is to predict the value of every query
node (upper row), according to its attribute. Each graph in the dataset has a different mapping from
attributes to values. We created a separate dataset for each k = {1, 2, 3, ...}, for which we trained a
different model, and measured per-node accuracy.
Although this is a contrived problem, it is relevant to any subgraph with keys that share more than
one query, and each query needs to attend to the keys differently. Such subgraphs are very common
in a variety of real-world domains. This problem tests the layer itself because it can be solved using a
single GNN layer, without suffering from multi-layer side-effects such as over-smoothing (Li et al.,
2018), over-squashing (Alon and Yahav, 2021), or vanishing gradients (Li et al., 2019). Our code
will be made publicly available, to serve as a testbed for future graph attention mechanisms.

Results Figure 3 shows the following surprising results: GAT with a single head (GAT1h ) failed to
fit the training set for any value of k, no matter for how many iterations it was trained, and after
trying various training methods. Thus, it expectedly fails to generalize (resulting in low test accuracy).
Using 8 heads, GAT8h successfully fits the training set, but generalizes poorly to the test set. In
contrast, GATv2 easily achieves 100% training and 100% test accuracies for any value of k, and even
for k=100 (not shown) and using a single head, thanks to its ability to perform dynamic attention.
These results clearly show the limitations of GAT, which are easily solved by GATv2. An additional
comparison to GIN, which could not fit this dataset, is provided in Figure 6 in Appendix D.1.

6
Published as a conference paper at ICLR 2022

Visualization Figure 1a (top) shows a heatmap of GAT’s attention scores in this D ICTIONARY-
L OOKUP problem. As shown, all query nodes q0 to q9 attend mostly to the eighth key (k8), and have
the same ranking of attention coefficients (Figure 1a (bottom)). In contrast, Figure 1b shows how
GATv2 can select a different key node for every query node, because it computes dynamic attention.
The role of multi-head attention Veličković et al. (2018) found the role of multi-head attention to
be stabilizing the learning process. Nevertheless, Figure 3 shows that increasing the number of heads
strictly increases training accuracy, and thus, the expressivity. Thus, GAT depends on having multiple
attention heads. In contrast, even a single GATv2 head generalizes better than a multi-head GAT.

4.2 ROBUSTNESS TO N OISE

We examine the robustness of dynamic and static attention to noise. In particular, we focus on
structural noise: given an input graph G = (V, E) and a noise ratio 0 ≤ p ≤ 1, we randomly sample
|E|×p non-existing edges E 0 from V×V\E. We then train the GNN on the noisy graph G 0 =(V, E ∪E 0 ).

72
GATv2 (this work) GATv2 (this work)
GAT 32 GAT
70
Accuracy

Accuracy

30
68

28
66

0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
p – noise ratio p – noise ratio
(a) ogbn-arxiv (b) ogbn-mag

Figure 4: Test accuracy compared to the noise ratio: GATv2 is more robust to structural noise
compared to GAT. Each point is an average of 10 runs, error bars show standard deviation.

Results Figure 9 shows the accuracy on two node-prediction datasets from the Open Graph Bench-
mark (OGB; Hu et al., 2020) as a function of the noise ratio p. As p increases, all models show
a natural decline in test accuracy in both datasets. Yet, thanks to their ability to compute dynamic
attention, GATv2 shows a milder degradation in accuracy compared to GAT, which shows a steeper
descent. We hypothesize that the ability to perform dynamic attention helps the models distinguishing
between given data edges (E) and noise edges (E 0 ); in contrast, GAT cannot distinguish between
edges, because it scores the source and target nodes separately. These results clearly demonstrate the
robustness of dynamic attention over static attention in noisy settings, which are common in reality.

4.3 P ROGRAMS : VAR M ISUSE

Setup VAR M ISUSE (Allamanis et al., 2018) is an inductive node-pointing problem that depends on
11 types of syntactic and semantic interactions between elements in computer programs.
We used the framework of Brockschmidt (2020), who performed an extensive hyperparameter tuning
by searching over 30 configurations for every GNN type. We took their best GAT hyperparameters
and used them to train GATv2, without further tuning.
Results As shown in Figure 5, GATv2 is more accurate than GAT and other GNNs in the SeenProj
test sets. Furthermore, GATv2 achieves an even higher improvement in the UnseenProj test set.
Overall, these results demonstrate the power of GATv2 in modeling complex relational problems,
especially since it outperforms extensively tuned models, without any further tuning by us.

7
Published as a conference paper at ICLR 2022

Figure 5: Accuracy (5 runs±stdev) on VAR M ISUSE. GATv2 is more accurate than all GNNs in both
test sets, using GAT’s hyperparameters. † previously reported by Brockschmidt (2020).

Model SeenProj UnseenProj

†
No- GCN 87.2±1.5 81.4±2.3
Attention GIN† 87.1±0.1 81.1±0.9
GAT† 86.9±0.7 81.2±0.9
Attention
GATv2 88.0±1.1 82.8±1.7

4.4 N ODE -P REDICTION

We further compare GATv2, GAT, and other GNNs on four node-prediction datasets from OGB.

Table 1: Average accuracy (Table 1a) and ROC-AUC (Table 1b) in node-prediction datasets (10
runs±std). In all datasets, GATv2 outperforms GAT. † – previously reported by Hu et al. (2020).

(a) (b)

Model Attn. Heads ogbn-arxiv ogbn-products ogbn-mag ogbn-proteins

†
GCN 0 71.74±0.29 78.97±0.33 30.43±0.25 72.51±0.35
GraphSAGE† 0 71.49±0.27 78.70±0.36 31.53±0.15 77.68±0.20
1 71.59±0.38 79.04±1.54 32.20±1.46 70.77±5.79
GAT
8 71.54±0.30 77.23±2.37 31.75±1.60 78.63±1.62
1 71.78±0.18 80.63±0.70 32.61±0.44 77.23±3.32
GATv2 (this work)
8 71.87±0.25 78.46±2.45 32.52±0.39 79.52±0.55

Results Results are shown in Table 1. In all settings and all datasets, GATv2 is more accurate than
GAT and the non-attentive GNNs. Interestingly, in the datasets of Table 1a, even a single head of
GATv2 outperforms GAT with 8 heads. In Table 1b (ogbn-proteins), increasing the number of heads
results in a major improvement for GAT (from 70.77 to 78.63), while GATv2 already gets most of
the benefit using a single attention head. These results demonstrate the superiority of GATv2 over
GAT in node prediction (and even with a single head), thanks to GATv2’s dynamic attention.

4.5 G RAPH -P REDICTION : QM9

Setup In the QM9 dataset (Ramakrishnan et al., 2014; Gilmer et al., 2017), each graph is a molecule
and the goal is to regress each graph to 13 real-valued quantum chemical properties. We used the
implementation of Brockschmidt (2020) who performed an extensive hyperparameter search over
500 configurations; we took their best-found configuration of GAT to implement GATv2.

Table 2: Average error rates (lower is better), 5 runs for each property, on the QM9 dataset. The best
result among GAT and GATv2 is marked in bold; the globally best result among all GNNs is marked
in bold and underline. † was previously tuned and reported by Brockschmidt (2020).

Predicted Property Rel. to

Model 1 2 3 4 5 6 7 8 9 10 11 12 13 GAT
GCN† 3.21 4.22 1.45 1.62 2.42 16.38 17.40 7.82 8.24 9.05 7.00 3.93 1.02 -1.5%
GIN† 2.64 4.67 1.42 1.50 2.27 15.63 12.93 5.88 18.71 5.62 5.38 3.53 1.05 -2.3%
GAT† 2.68 4.65 1.48 1.53 2.31 52.39 14.87 7.61 6.86 7.64 6.54 4.11 1.48 +0%
GATv2 2.65 4.28 1.41 1.47 2.29 16.37 14.03 6.07 6.28 6.60 5.97 3.57 1.59 -11.5%

Results Table 2 shows the main results: GATv2 achieves a lower (better) average error than GAT, by
11.5% relatively. GAT achieves the overall highest average error. In some properties, the non-attentive

8
Published as a conference paper at ICLR 2022

GNNs, GCN and GIN, perform best. We hypothesize that attention is not needed in modeling these
properties. Generally, GATv2 achieves the lowest overall average relative error (rightmost column).

4.6 L INK -P REDICTION

We compare GATv2, GAT, and other GNNs in link-prediction datasets from OGB.

Table 3: Average Hits@50 (Table 3a) and mean reciprocal rank (MRR) (Table 3b) in link-prediction
benchmarks from OGB (10 runs±std). The best result among GAT and GATv2 is marked in bold;
the best result among all GNNs is marked in bold and underline. † was reported by Hu et al. (2020).

(a) (b)

ogbl-collab ogbl-citation2
Model Attn. Heads w/o val edges w/ val edges
No- GCN† 44.75±1.07 47.14±1.45 80.04±0.25
Attention GraphSAGE† 48.10±0.81 54.63±1.12 80.44±0.10
GAT1h 39.32±3.26 48.10±4.80 79.84±0.19
GAT
GAT8h 42.37±2.99 46.63±2.80 75.95±1.31
GATv2
GATv21h 42.00±2.40 48.02±2.77 80.33±0.13
GATv28h 42.85±2.64 49.70±3.08 80.14±0.71

Results Table 3 shows that in all datasets, GATv2 achieves a higher MRR than GAT, which achieves
the lowest MRR. However, the non-attentive GraphSAGE performs better than all attentive GNNs.
We hypothesize that attention might not be needed in these datasets. Another possibility is that
dynamic attention is especially useful in graphs that have high node degrees: in ogbn-products and
ogbn-proteins (Table 1) the average node degrees are 50.5 and 597, respectively (see Table 5 in
Appendix C). ogbl-collab and ogbl-citation2 (Table 3), however, have much lower average node
degrees – of 8.2 and 20.7. We hypothesize that a dynamic attention mechanism is especially useful to
select the most relevant neighbors when the total number of neighbors is high. We leave the study of
the effect of the datasets’s average node degrees on the optimal GNN architecture for future work.

4.7 D ISCUSSION

In all examined benchmarks, we found that GATv2 is more accurate than GAT. Further, we found
that GATv2 is significantly more robust to noise than GAT. In the synthetic D ICTIONARY L OOKUP
benchmark (Section 4.1), GAT fails to express the data, and thus achieves even poor training accuracy.
In few of the benchmarks (Table 3 and some of the properties in Table 2) – a non-attentive model
such as GCN or GIN achieved a higher accuracy than all GNNs that do use attention.

Which graph attention mechanism should I use? It is usually impossible to determine in advance
which architecture would perform best. A theoretically weaker model may perform better in practice,
because a stronger model might overfit the training data if the task is “too simple” and does not
require such expressiveness. Intuitively, we believe that the more complex the interactions between
nodes are – the more benefit a GNN can take from theoretically stronger graph attention mechanisms
such as GATv2. The main question is whether the problem has a global ranking of “influential” nodes
(GAT is sufficient), or do different nodes have different rankings of neighbors (use GATv2).
Veličković, the author of GAT, has confirmed on Twitter 6 that GAT was designed to work in the
“easy-to-overfit” datasets of the time (2017), such as Cora, Citeseer and Pubmed (Sen et al., 2008),
where the data might had an underlying static ranking of “globally important” nodes. Veličković
agreed that newer and more challenging benchmarks may demand stronger attention mechanisms
such as GATv2. In this paper, we revisit the traditional assumptions and show that many modern graph
benchmarks and datasets contain more complex interactions, and thus require dynamic attention.

6
https://twitter.com/PetarV_93/status/1399685979506675714

9
Published as a conference paper at ICLR 2022

5 R ELATED W ORK

Attention in GNNs Modeling pairwise interactions between elements in graph-structured data goes
back to interaction networks (Battaglia et al., 2016; Hoshen, 2017) and relational networks (Santoro
et al., 2017). The GAT formulation of Veličković et al. (2018) rose as the most popular framework
for attentional GNNs, thanks to its simplicity, generality, and applicability beyond reinforcement
learning (Denil et al., 2017; Duan et al., 2017). Nevertheless, in this work, we show that the popular
and widespread definition of GAT is severely constrained to static attention only.
Other graph attention mechanisms Many works employed GNNs with attention mechanisms
other than the standard GAT’s (Zhang et al., 2018; Thekumparampil et al., 2018; Gao and Ji, 2019;
Lukovnikov and Fischer, 2021; Shi et al., 2020; Dwivedi and Bresson, 2020; Busbridge et al., 2019;
Rong et al., 2020a; Veličković et al., 2020), and Lee et al. (2018) conducted an extensive survey
of attention types in GNNs. However, none of these works identified the monotonicity of GAT’s
attention mechanism, the theoretical differences between attention types, nor empirically compared
their performance. Kim and Oh (2021) compared two graph attention mechanisms empirically, but in a
specific self-supervised scenario, without observing the theoretical difference in their expressiveness.

The static attention of GAT Qiu et al. (2018) recognized the order-preserving property of GAT, but
did not identify the severe theoretical constraint that this property implies: the inability to perform
dynamic attention (Theorem 1). Furthermore, they presented GAT’s monotonicity as a desired trait (!)
To the best of our knowledge, our work is the first work to recognize the inability of GAT to perform
dynamic attention and its practical harmful consequences.

6 C ONCLUSION

In this paper, we identify that the popular and widespread Graph Attention Network does not compute
dynamic attention. Instead, the attention mechanism in the standard definition and implementations
of GAT is only static: for any query, its neighbor-scoring is monotonic with respect to per-node
scores. As a result, GAT cannot even express simple alignment problems. To address this limitation,
we introduce a simple fix and propose GATv2: by modifying the order of operations in GAT, GATv2
achieves a universal approximator attention function and is thus strictly more powerful than GAT.
We demonstrate the empirical advantage of GATv2 over GAT in a synthetic problem that requires dy-
namic selection of nodes, and in 11 benchmarks from OGB and other public datasets. Our experiments
show that GATv2 outperforms GAT in all benchmarks while having the same parametric cost.
We encourage the community to use GATv2 instead of GAT whenever comparing new GNN ar-
chitectures to the common strong baselines. In complex tasks and domains and in challenging
datasets, a model that uses GAT as an internal component can replace it with GATv2 to bene-
fit from a strictly more powerful model. To this end, we make our code publicly available at
https://github.com/tech-srl/how_attentive_are_gats , and GATv2 is available
as part of the PyTorch Geometric library, the Deep Graph Library, and TensorFlow GNN. An anno-
tated implementation is available at https://nn.labml.ai/graphs/gatv2/ .

ACKNOWLEDGMENTS

We thank Gail Weiss for the helpful discussions, thorough feedback, and inspirational paper (Weiss
et al., 2018). We also thank Petar Veličković for the useful discussion about the complexity and
implementation of GAT.

R EFERENCES
Miltiadis Allamanis, Marc Brockschmidt, and Mahmoud Khademi. Learning to represent programs with graphs.
In International Conference on Learning Representations, 2018. URL https://openreview.net/
forum?id=BJOFETxR-.

10
Published as a conference paper at ICLR 2022

Uri Alon and Eran Yahav. On the bottleneck of graph neural networks and its practical implications. In Interna-
tional Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=
i80OPhOCVH2.

James Atwood and Don Towsley. Diffusion-convolutional neural networks. In Advances in neural information
processing systems, pages 1993–2001, 2016.

Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to
align and translate. CoRR, abs/1409.0473, 2014. URL http://arxiv.org/abs/1409.0473.

Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, and Koray kavukcuoglu. Interaction
networks for learning about objects, relations and physics. In Proceedings of the 30th International Conference
on Neural Information Processing Systems, pages 4509–4517, 2016.

Marc Brockschmidt. Gnn-film: Graph neural networks with feature-wise linear modulation. Proceedings of
the 36th International Conference on Machine Learning, ICML, 2020. URL https://github.com/
microsoft/tf-gnn-samples.

Michael M Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Vandergheynst. Geometric deep
learning: going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18–42, 2017.

Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Veličković. Geometric deep learning: Grids, groups,
graphs, geodesics, and gauges, 2021.

Dan Busbridge, Dane Sherburn, Pietro Cavallo, and Nils Y Hammerla. Relational graph attention networks.
arXiv preprint arXiv:1904.05811, 2019.

George Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of control, signals

and systems, 2(4):303–314, 1989.

Misha Denil, Sergio Gómez Colmenarejo, Serkan Cabi, David Saxton, and Nando de Freitas. Programmable
agents. arXiv preprint arXiv:1706.06383, 2017.

Yan Duan, Marcin Andrychowicz, Bradly Stadie, Jonathan Ho, Jonas Schneider, Ilya Sutskever, Pieter Abbeel,
and Wojciech Zaremba. One-shot imitation learning. In Proceedings of the 31st International Conference on
Neural Information Processing Systems, pages 1087–1098, 2017.

David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alán Aspuru-
Guzik, and Ryan P Adams. Convolutional networks on graphs for learning molecular fingerprints. In
Advances in neural information processing systems, pages 2224–2232, 2015.

Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs. arXiv preprint
arXiv:2012.09699, 2020.

Vijay Prakash Dwivedi, Chaitanya K Joshi, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Benchmarking
graph neural networks. arXiv preprint arXiv:2003.00982, 2020.

Matthias Fey and Jan E. Lenssen. Fast graph representation learning with PyTorch Geometric. In ICLR Workshop
on Representation Learning on Graphs and Manifolds, 2019.

Ken-Ichi Funahashi. On the approximate realization of continuous mappings by neural networks. Neural
networks, 2(3):183–192, 1989.

Hongyang Gao and Shuiwang Ji. Graph representation learning via hard and channel-wise attention networks.
In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,
pages 741–749, 2019.

Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing
for quantum chemistry. In Proceedings of the 34th International Conference on Machine Learning-Volume
70, pages 1263–1272. JMLR. org, 2017.

Aleksa Gordić. pytorch-gat. https://github.com/gordicaleksa/pytorch-GAT, 2020.

Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. In
Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., volume 2, pages 729–
734. IEEE, 2005.

Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. In Advances
in neural information processing systems, pages 1024–1034, 2017.

11
Published as a conference paper at ICLR 2022

Kurt Hornik. Approximation capabilities of multilayer feedforward networks. Neural networks, 4(2):251–257,
1991.

Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal
approximators. Neural networks, 2(5):359–366, 1989.

Yedid Hoshen. Vain: attentional multi-agent predictive modeling. In Proceedings of the 31st International
Conference on Neural Information Processing Systems, pages 2698–2708, 2017.

Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure
Leskovec. Open graph benchmark: Datasets for machine learning on graphs. arXiv preprint arXiv:2005.00687,
2020.

Binxuan Huang and Kathleen M Carley. Syntax-aware aspect level sentiment classification with graph attention
networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and
the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5472–5480,
2019.

Qian Huang, Horace He, Abhay Singh, Ser-Nam Lim, and Austin Benson. Combining label propagation and
simple models out-performs graph neural networks. In International Conference on Learning Representations,
2021. URL https://openreview.net/forum?id=8E1-f3VhX1o.

Chaitanya Joshi. Transformers are graph neural networks. The Gradient, 2020.

Dongkwan Kim and Alice Oh. How to find your friendly neighborhood: Graph attention design with
self-supervision. In International Conference on Learning Representations, 2021. URL https://
openreview.net/forum?id=Wi5KUNlqWty.

Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR,
2017.

Vineet Kosaraju, Amir Sadeghian, Roberto Martín-Martín, Ian Reid, Hamid Rezatofighi, and Silvio Savarese.
Social-bigat: Multimodal trajectory forecasting using bicycle-gan and graph attention networks. In H. Wallach,
H. Larochelle, A. Beygelzimer, F. d Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Infor-
mation Processing Systems, volume 32. Curran Associates, Inc., 2019. URL https://proceedings.
neurips.cc/paper/2019/file/d09bf41544a3365a46c9077ebb5e35c3-Paper.pdf.

John Boaz Lee, Ryan A Rossi, Sungchul Kim, Nesreen K Ahmed, and Eunyee Koh. Attention models in graphs:
A survey. arXiv preprint arXiv:1807.07984, 2018.

Moshe Leshno, Vladimir Ya Lin, Allan Pinkus, and Shimon Schocken. Multilayer feedforward networks with a
nonpolynomial activation function can approximate any function. Neural networks, 6(6):861–867, 1993.

Guohao Li, Matthias Muller, Ali Thabet, and Bernard Ghanem. Deepgcns: Can gcns go as deep as cnns? In
Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9267–9276, 2019.

Qimai Li, Zhichao Han, and Xiao-Ming Wu. Deeper insights into graph convolutional networks for semi-
supervised learning. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018.

Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. In
International Conference on Learning Representations, 2016.

Denis Lukovnikov and Asja Fischer. Gated relational graph attention networks, 2021. URL https://
openreview.net/forum?id=v-9E8egy_i.

Thang Luong, Hieu Pham, and Christopher D. Manning. Effective approaches to attention-based neural
machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language
Processing, EMNLP 2015, Lisbon, Portugal, September 17-21, 2015, pages 1412–1421, 2015. URL http:
//aclweb.org/anthology/D/D15/D15-1166.pdf.

Nianzu Ma, Sahisnu Mazumder, Hao Wang, and Bing Liu. Entity-aware dependency-based deep graph attention
network for comparative preference classification. In Proceedings of the 58th Annual Meeting of the
Association for Computational Linguistics, pages 5782–5788, 2020.

Federico Monti, Davide Boscaini, Jonathan Masci, Emanuele Rodola, Jan Svoboda, and Michael M Bronstein.
Geometric deep learning on graphs and manifolds using mixture model cnns. In Proceedings of the IEEE
conference on computer vision and pattern recognition, pages 5115–5124, 2017.

12
Published as a conference paper at ICLR 2022

Deepak Nathani, Jatin Chauhan, Charu Sharma, and Manohar Kaul. Learning attention-based embeddings for
relation prediction in knowledge graphs. In Proceedings of the 57th Annual Meeting of the Association for
Computational Linguistics, pages 4710–4723, 2019.
Sejun Park, Chulhee Yun, Jaeho Lee, and Jinwoo Shin. Minimum width for universal approximation. In Interna-
tional Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=
O-XJwyoIF-k.
Allan Pinkus. Approximation theory of the mlp model. Acta Numerica 1999: Volume 8, 8:143–195, 1999.
Jiezhong Qiu, Jian Tang, Hao Ma, Yuxiao Dong, Kuansan Wang, and Jie Tang. Deepinf: Social influence
prediction with deep learning. In Proceedings of the 24th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining (KDD’18), 2018.
Raghunathan Ramakrishnan, Pavlo O Dral, Matthias Rupp, and O Anatole Von Lilienfeld. Quantum chemistry
structures and properties of 134 kilo molecules. Scientific data, 1:140022, 2014.
Yu Rong, Yatao Bian, Tingyang Xu, Weiyang Xie, Ying Wei, Wenbing Huang, and Junzhou Huang. Self-
supervised graph transformer on large-scale molecular data. Advances in Neural Information Processing
Systems, 33, 2020a.
Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. Dropedge: Towards deep graph convolutional
networks on node classification. In International Conference on Learning Representations, 2020b. URL
https://openreview.net/forum?id=Hkx1qkrKPr.
Adam Santoro, David Raposo, David GT Barrett, Mateusz Malinowski, Razvan Pascanu, Peter Battaglia, and
Timothy Lillicrap. A simple neural network module for relational reasoning. In Proceedings of the 31st
International Conference on Neural Information Processing Systems, pages 4974–4983, 2017.
Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph
neural network model. IEEE Transactions on Neural Networks, 20(1):61–80, 2008.
Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi-Rad. Collective
classification in network data. AI magazine, 29(3):93–93, 2008.
Yunsheng Shi, Zhengjie Huang, Shikun Feng, and Yu Sun. Masked label prediction: Unified massage passing
model for semi-supervised classification. arXiv preprint arXiv:2009.03509, 2020.
Kiran K Thekumparampil, Chong Wang, Sewoong Oh, and Li-Jia Li. Attention-based graph neural network for
semi-supervised learning. arXiv preprint arXiv:1803.03735, 2018.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser,
and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages
6000–6010, 2017.
Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph
attention networks. In International Conference on Learning Representations, 2018.
Petar Veličković, Lars Buesing, Matthew Overlan, Razvan Pascanu, Oriol Vinyals, and Charles Blundell. Pointer
graph networks. Advances in Neural Information Processing Systems, 33, 2020.
Petar et al. Veličković. Graph attention networks. 2018.
Guangtao Wang, Rex Ying, Jing Huang, and Jure Leskovec. Improving graph attention networks with large
margin-based constraints. arXiv preprint arXiv:1910.11945, 2019a.
Minjie Wang, Da Zheng, Zihao Ye, Quan Gan, Mufei Li, Xiang Song, Jinjing Zhou, Chao Ma, Lingfan Yu,
Yu Gai, Tianjun Xiao, Tong He, George Karypis, Jinyang Li, and Zheng Zhang. Deep graph library: A
graph-centric, highly-performant package for graph neural networks. arXiv preprint arXiv:1909.01315,
2019b.
Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu. Heterogeneous graph
attention network. In The World Wide Web Conference, pages 2022–2032, 2019c.
Yangkun Wang. Bag of tricks of semi-supervised classification with graph neural networks. arXiv preprint
arXiv:2103.13355, 2021.
Gail Weiss, Yoav Goldberg, and Eran Yahav. On the practical computational power of finite precision rnns
for language recognition. In Proceedings of the 56th Annual Meeting of the Association for Computational
Linguistics (Volume 2: Short Papers), pages 740–745, 2018.

13
Published as a conference paper at ICLR 2022

Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. Simplifying graph
convolutional networks. In International conference on machine learning, pages 6861–6871. PMLR, 2019.

Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive
survey on graph neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2020.

Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks?
In International Conference on Learning Representations, 2019. URL https://openreview.net/
forum?id=ryGs6iA5Km.

Yiding Yang, Jiayan Qiu, Mingli Song, Dacheng Tao, and Xinchao Wang. Distilling knowledge from graph
convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition (CVPR), June 2020.
Hanqing Zeng, Hongkuan Zhou, Ajitesh Srivastava, Rajgopal Kannan, and Viktor Prasanna. Graphsaint: Graph
sampling based inductive learning method. arXiv preprint arXiv:1907.04931, 2019.

Jiani Zhang, Xingjian Shi, Junyuan Xie, Hao Ma, Irwin King, and Dit-Yan Yeung. Gaan: Gated attention
networks for learning on large and spatiotemporal graphs. In Proceedings of the Thirty-Fourth Conference on
Uncertainty in Artificial Intelligence, pages 339–349, 2018.

Kai Zhang, Yaokang Zhu, Jun Wang, and Jie Zhang. Adaptive structural fingerprints for graph attention networks.
In International Conference on Learning Representations, 2020. URL https://openreview.net/
forum?id=BJxWx0NYPr.

Lingxiao Zhao and Leman Akoglu. Pairnorm: Tackling oversmoothing in gnns. In International Conference on
Learning Representations, 2020. URL https://openreview.net/forum?id=rkecl1rtwB.

14
Published as a conference paper at ICLR 2022

A P ROOF FOR T HEOREM 2

For brevity, we repeat our definition of dynamic attention (Definition 3.2):

Definition 3.2
(Dynamic attention). A (possibly infinite) family of scoring functions F ⊆

Rd × Rd → R computes dynamic scoring for a given set of key vectors K = {k1 , ..., kn } ⊂ Rd
and query vectors Q = {q1 , ..., qm } ⊂ Rd , if for any mapping ϕ:
[m] → [n] there exists f ∈ F such
that for any query i ∈ [m] and any key j6=ϕ(i) ∈ [n]: f qi , kϕ(i) > f (qi , kj ). We say that a family
of attention functions computes dynamic attention for K and Q, if its scoring function computes
dynamic scoring, possibly followed by monotonic normalization such as softmax.
Theorem 2. A GATv2 layer computes dynamic attention for any set of node representations K =
Q = {h1 , ..., hn } .

Proof. Let G = (V, E) be a graph modeled by a GATv2 layer, having node representations
{h1 , ..., hn }, and let ϕ : [n] → [n] be any node mapping [n] → [n]. We define g : R2d → R
as follows:   
1 ∃i : x = hi khϕ(i)
g (x) = (8)
0 otherwise

Next, we define a continues function ge : R2d → R that equals to g in only specific n2 inputs:
ge([hi khj ]) = g([hi khj ]), ∀i, j ∈ [n] (9)
For all other inputs x ∈ R2d , ge(x) realizes to any values that maintain the continuity of ge (this is
possible because we fixed the values of ge for only a finite set of points). 7
Thus, for every node i ∈ V and j6=ϕ(i) ∈ V:
 

1 = ge hi khϕ(i) > ge ([hi khj ]) = 0 (10)

If we concatenate the two input vectors, and define the scoring function e of GATv2 (Equation (7)) as
a function of the concatenated vector [hi khj ], from the universal approximation theorem (Hornik
et al., 1989; Cybenko, 1989; Funahashi, 1989; Hornik, 1991), e can approximate ge for any compact
subset of R2d .
Thus, for any sufficiently small  (any 0 <  < 1/2) there exist parameters W and a such that for
every node i ∈ V and every j6=ϕ(i) :


eW ,a hi , hϕ(i) > 1 −  > 0 +  > eW ,a (hi , hj ) (11)
and due to the increasing monotonicity of softmax:
αi,ϕ(i) > αi,j (12)

The choice of nonlinearity In general, these results hold if GATv2 had used any common non-
polynomial activation function (such as ReLU, sigmoid, or the hyperbolic tangent function). The
LeakyReLU activation function of GATv2 does not change its universal approximation ability (Leshno
et al., 1993; Pinkus, 1999; Park et al., 2021), and it was chosen only for consistency with the original
definition of GAT.
7
The function ge is a function that we define for the ease of proof, because the universal approximation
theorem is defined for continuous functions, and we only need the scoring function of GATv2 e to approximate
the mapping ϕ in a finite set of points. So, we need the attention function e to approximate g (from Equation 8)
in some specific points. But, since g is not continuous, we define ge and use the universal approximation theorem
for ge. Since e approximates ge, e also approximates g in our specific points, as a special case. We only require
that ge will be identical to g in specific n2 points {[hi khj ] | i, j ∈ [n]}. For the rest of the input space, we don’t
have any requirement on the value of ge, except for maintaining the continuity of ge. There exist infinitely many
such possible ge for every given set of keys, queries and a mapping ϕ, but the concrete functions are not needed
for the proof.

15
Published as a conference paper at ICLR 2022

B T RAINING DETAILS

In this section we elaborate on the training details of all of our experiments. All models use residual
connections as in Veličković et al. (2018). All used code and data are publicly available under the
MIT license.

B.1 N ODE - AND L INK -P REDICTION

We used the provided splits of OGB (Hu et al., 2020) and the Adam optimizer. We tuned the
following hyperparameters: number of layers ∈ {2, 3, 6}, hidden size ∈ {64, 128, 256}, learning rate
∈ {0.0005, 0.001, 0.005, 0.01} and sampling method – full batch, GraphSAINT (Zeng et al., 2019)
and NeighborSampling (Hamilton et al., 2017). We tuned hyperparameters according to validation
score and early stopping. The final hyperparameters are detailed in Table 4.

Dataset # layers Hidden size Learning rate Sampling method

ogbn-arxiv 3 256 0.01 GraphSAINT
ogbn-products 3 128 0.001 NeighborSampling
ogbn-mag 2 256 0.01 NeighborSampling
ogbn-proteins 6 64 0.01 NeighborSampling
ogbl-collab 3 64 0.001 Full Batch
ogbl-citation2 3 256 0.0005 NeighborSampling

Table 4: Training details of node- and link-prediction datasets.

B.2 ROBUSTNESS TO N OISE

In these experiments, we used the same best-found hyperparameters in node-prediction, with 8

attention heads in ogbn-arxiv and 1 head in ogbn-mag. Each point is an average of 10 runs.

B.3 S YNTHETIC B ENCHMARK : D ICTIONARY L OOKUP

In all experiments, we used a learning rate decay of 0.5, a hidden size of d = 128, a batch size of
1024, and the Adam optimizer.
We created a separate dataset for every graph size (k), and we split each such dataset to train and
test with a ratio of 80:20. Since this is a contrived problem, we did not use a validation set, and the
reported test results can be thought of as validation results. Every model was trained on a fixed value
of k. Every key node (bottom row in Figure 2) was encoded as a sum of learned attribute embedding
and a value embedding, followed by ReLU.
We experimented with layer normalization, batch normalization, dropout, various activation functions
and various learning rates. None of these changed the general trend, so the experiments in Figure 3
were conducted without any normalization, without dropout and a learning rate of 0.001.

B.4 P ROGRAMS : VAR M ISUSE

We used the code, splits, and the same best-found configurations as Brockschmidt (2020), who
performed an extensive hyperparameter tuning by searching over 30 configurations for each GNN
type. We trained each model five times.
We took the best-found hyperparameters of Brockschmidt (2020) for GAT and used them to train
GATv2, without any further tuning.

B.5 G RAPH -P REDICTION : QM9

We used the code and splits of Brockschmidt (2020) who performed an extensive hyperparameter
search over 500 configurations. We took the best-found hyperparameters of Brockschmidt (2020)

16
Published as a conference paper at ICLR 2022

for GAT and used them to train GATv2. The only minor change from GAT is placing a residual
connection after every layer, rather than after every other layer, which is within the experimented
hyperparameter search that was reported by Brockschmidt (2020).

B.6 C OMPUTE AND R ESOURCES

Our experiments consumed approximately 100 days of GPU in total. We used cloud GPUs of type
V100, and we used RTX 3080 and 3090 in local GPU machines.

C DATA S TATISTICS

C.1 N ODE - AND L INK -P REDICTION DATASETS

Statistics of the OGB datasets we used for node- and link-prediction are shown in Table 5.

Dataset # nodes # edges Avg. node degree Diameter

ogbn-arxiv 169,343 1,166,243 13.7 23
ogbn-mag 1,939,743 21,111,007 21.7 6
ogbn-products 2,449,029 61,859,140 50.5 27
ogbn-proteins 132,534 39,561,252 597.0 9
ogbl-collab 235,868 1,285,465 8.2 22
ogbl-citation2 2,927,963 30,561,187 20.7 21

Table 5: Statistics of the OGB datasets (Hu et al., 2020).

C.2 QM9

Statistics of the QM9 dataset, as used in Brockschmidt (2020) are shown in Table 6.

Training Validation Test

# examples 110,462 10,000 10,000
# nodes - average 18.03 18.06 18.09
# edges - average 18.65 18.67 18.72
Diameter - average 6.35 6.35 6.35

Table 6: Statistics of the QM9 chemical dataset (Ramakrishnan et al., 2014) as used by Brockschmidt
(2020).

C.3 VAR M ISUSE

Statistics of the VAR M ISUSE dataset, as used in Allamanis et al. (2018) and Brockschmidt (2020),
are shown in Table 7.

Training Validation UnseenProject Test SeenProject Test

# graphs 254360 42654 117036 59974
# nodes - average 2377 1742 1959 3986
# edges - average 7298 7851 5882 12925
Diameter - average 7.88 7.88 7.78 7.82

Table 7: Statistics of the VAR M ISUSE dataset (Allamanis et al., 2018) as used by Brockschmidt
(2020).

17
Published as a conference paper at ICLR 2022

100
90
80
70
Accuracy 60
50
GATv21h train
40 GATv21h test
30 GIN train
20 GIN test
10
0
10 20 30 40 50 60 70

k (number of different keys in each graph)

Figure 6: Train and test accuracy across graph sizes in the D ICTIONARY L OOKUP problem. GATv2
easily achieves 100% train and test accuracy even for k=100 and using only a single head. GIN
(Xu et al., 2019), although considered as more expressive than other GNNs, cannot perfectly fit the
training data (with a model size of d = 128) starting from k=20.

D A DDITIONAL R ESULTS

D.1 D ICTIONARY L OOKUP

Figure 6 shows additional comparison between GATv2 and GIN (Xu et al., 2019) in the D ICTIO -
NARY L OOKUP problem. GATv2 easily achieves 100% train and test accuracy even for k=100 and
using only a single head. GIN, although considered as more expressive than other GNNs, cannot
perfectly fit the training data (with a model size of d = 128) starting from k=20.

D.2 QM9

Standard deviation for the QM9 results of Section 4.5 are presented in Table 8.

Predicted Property
Model 1 2 3 4 5 6 7
†
GCN 3.21±0.06 4.22±0.45 1.45±0.01 1.62±0.04 2.42±0.14 16.38±0.49 17.40±3.56
GIN† 2.64±0.11 4.67±0.52 1.42±0.01 1.50±0.09 2.27±0.09 15.63±1.40 12.93±1.81
GAT1h 3.08±0.08 7.82±1.42 1.79±0.10 3.96±1.51 3.58±1.03 35.43±29.9 116.5±10.65
GAT8h † 2.68±0.06 4.65±0.44 1.48±0.03 1.53±0.07 2.31±0.06 52.39±42.58 14.87±2.88
GATv21h 3.04±0.06 6.38±0.62 1.68±0.04 2.18±0.61 2.82±0.25 20.56±0.70 77.13±37.93
GATv28h 2.65±0.05 4.28±0.27 1.41±0.04 1.47±0.03 2.29±0.15 16.37±0.97 14.03±1.39

Predicted Property Rel. to

Model 8 9 10 11 12 13 GAT8h
GCN† 7.82±0.80 8.24±1.25 9.05±1.21 7.00±1.51 3.93±0.48 1.02±0.05 -1.5%
GIN† 5.88±1.01 18.71±23.36 5.62±0.81 5.38±0.75 3.53±0.37 1.05±0.11 -2.3%
GAT1h 28.10±16.45 20.80±13.40 15.80±5.87 10.80±2.18 5.37±0.26 3.11±0.14 +134.1%
GAT8h † 7.61±0.46 6.86±0.53 7.64±0.92 6.54±0.36 4.11±0.27 1.48±0.87 +0%
GATv21h 10.19±0.63 22.56±17.46 15.04±4.58 22.94±17.34 5.23±0.36 2.46±0.65 +91.6%
GATv28h 6.07±0.77 6.28±0.83 6.60±0.79 5.97±0.94 3.57±0.36 1.59±0.96 -11.5%

Table 8: Average error rates (lower is better), 5 runs ± standard deviation for each property, on
the QM9 dataset. The best result among GAT and GATv2 is marked in bold; the globally best
result among all GNNs is marked in bold and underline. † was previously tuned and reported by
Brockschmidt (2020).

18
Published as a conference paper at ICLR 2022

D.3 P UBMED C ITATION N ETWORK

We tuned the following parameters for both GAT and GATv2: number of layers ∈ {0, 1, 2}, hidden
size ∈ {8, 16, 32}, number of heads ∈ {1, 4, 8}, dropout ∈ {0.4, 0.6, 0.8}, bias ∈ {T rue, F alse},
share weights ∈ {T rue, F alse}, use residual ∈ {T rue, F alse}. Table 9 shows the test accuracy
(100 runs±stdev) using the best hyperparameters found for each model.

Table 9: Accuracy (100 runs±stdev) on Pubmed. GATv2 is more accurate than GAT.

Model Accuracy
GAT 78.1±0.59
GATv2 78.5±0.38

It is important to note that PubMed has only 60 training nodes, which hinders expressive models
such as GATv2 from exploiting their approximation and generalization advantages. Still, GATv2
is more accurate than GAT even in this small dataset. In Table 14, we show that this difference is
statistically significant (p-value < 0.0001).

E A DDITIONAL C OMPARISON WITH T RANSFORMER - STYLE ATTENTION

(DPGAT)

The main goal of our paper is to highlight a severe theoretical limitation of the highly popular GAT
architecture, and propose a minimal fix.
We perform additional empirical comparison to DPGAT, which follows Luong et al. (2015) and the
dot-product attention of the Transformer (Vaswani et al., 2017). We define DPGAT as:

>  p
e (hi , hj ) = h> >


DPGAT (Vaswani et al., 2017): i Q · hj K / dk (13)

Variants of DPGAT were used in prior work (Gao and Ji, 2019; Dwivedi and Bresson, 2020; Rong
et al., 2020a; Veličković et al., 2020; Kim and Oh, 2021), and we consider it here for the conceptual
and empirical comparison with GAT.
Despite its popularity, DPGAT is strictly weaker than GATv2. DPGAT provably performs dynamic
attention for any set of node representations only if they are linearly independent (see Theorem 3
and its proof in Appendix E.1). Otherwise, there are examples of node representations that are
linearly dependent and mappings ϕ, for which dynamic attention does not hold (Appendix E.2).
This constraint is not harmful when violated in practice, because every node has only a small set of
neighbors, rather than all possible nodes in the graph; further, some nodes possibly never need to be
“selected” in practice.

E.1 P ROOF THAT DPGAT P ERFORMS DYNAMIC ATTENTION FOR L INEARLY I NDEPENDENT
N ODE R EPRESENTATIONS

Theorem 3. A DPGAT layer computes dynamic attention for any set of node representations K =
Q = {h1 , ..., hn } that are linearly independent.

Proof. Let G = (V, E) be a graph modeled by a DPGAT layer, having linearly independent node
representations {h1 , ..., hn }. Let ϕ : [n] → [n] be any node mapping [n] → [n].
We denote the ith row of a matrix M as Mi .
We define a matrix P as: 
1 j = ϕ(i)
Pi,j = (14)
0 otherwise

Let X ∈ Rn × Rd be the matrix holding the graph’s node representations as its rows:

19
Published as a conference paper at ICLR 2022

— h1 —
 
— h2 —
X= ..  (15)
 . 
— hn —

Since the rows of X are linearly independent, it necessarily holds that d ≥ n.

Next, we find weight matrices Q ∈ Rd × Rd and K ∈ Rd × Rd such that:
(XQ) · (XK)> = P (16)
To satisfy Equation (16), we choose Q and K such that XQ = U and XK = P > U where U is
an orthonormal matrix (U · U > = U > · U = I).
We can obtain U using the singular value decomposition (SVD) of X:
X = U ΣV > (17)
Since Σ ∈ Rn × Rn and X has a full rank, Σ is invertible, and thus:
XV Σ−1 = U (18)
Now, we define Q as follows:
Q = V Σ−1 (19)
Note that XQ = U , as desired.
To find K that satisfies XK = P > U , we use Equation (17) and require:
U ΣV > K = P > U (20)
and thus:
K = V Σ−1 U T P > U (21)

We define: p
z (hi , hj ) = e (hi , hj ) · dk (22)
Where e is the attention score function of DPGAT (Equation (13)).
Now, for a query i and a key j, and the corresponding representations hi , hj :

>
z (hi , hj ) = h> >


i Q · hj K (23)
>
= (Xi Q) · (Xj K) (24)

Since Xi Q = (XQ)i and Xj K = (XK)j , we get

 >
z (hi , hj ) = (XQ)i · (XK)j = Pi,j (25)

Therefore: 
1 j = ϕ(i)
z (hi , hj ) = (26)
0 otherwise

And thus:  √
1/ dk j = ϕ(i)
e (hi , hj ) = (27)
0 otherwise

To conclude, for every selected query i and any key j6=ϕ(i) :

e hi , hϕ(i) > e (hi , hj ) (28)
and due to the increasing monotonicity of softmax:
αi,ϕ(i) > αi,j (29)

20
Published as a conference paper at ICLR 2022

Hence, a DPGAT layer computes dynamic attention.

In the case that d > n, we apply SVD to the full-rank matrix XX > ∈ Rn×n , and follow the same
steps to construct Q and K.
In the case that Q ∈ Rd × Rdk and K ∈ Rd × Rdk and dk > d, we can use the same Q and K
(Equations (19) and (21)) padded with zeros. We define the Q0 ∈ Rd × Rdkey and K 0 ∈ Rd × Rdkey
as follows:

0 Qi,j j ≤ d
Qi,j = (30)
0 otherwise

0 Ki,j j ≤ d
Ki,j = (31)
0 otherwise

E.2 DPGAT IS STRICTLY WEAKER THAN GAT V 2

There are examples of node representations that are linearly dependent and mappings ϕ, for which
dynamic attention does not hold. First, we show a simple 2-dimensional example, and then we show
the general case of such examples.

y
h2 = x̂ + 2ŷ

h1 = x̂ + ŷ

h0 = x̂
x

Figure 7: An example for node representations that are linearly dependent, for which DPGAT cannot
compute dynamic attention, because no query vector q ∈ R2 can “select” h1 .

Consider the following linearly dependent set of vectors K = Q (Figure 7):

h0 = x̂ (32)
h1 = x̂ + ŷ (33)
h2 = x̂ + 2ŷ (34)
where x̂ and ŷ are the cartesian unit vectors. We define β ∈ {0, 1, 2} to express {h0 , h1 , h2 } using
the same expression:
hβ = x̂ + β ŷ (35)
Let q ∈ Q be any query vector. For brevity, we define the unscaled dot-product attention score as s:
p
s (q, hβ ) = e (q, hβ ) · dk (36)
Where e is the attention score function of DPGAT (Equation (13)). The (unscaled) attention score
between q and {h0 , h1 , h2 } is:

>
s (q, hβ ) = q > Q h>


βK (37)
  >
>
= q > Q (x̂ + β ŷ) K


(38)
>
= q > Q x̂> K + β ŷ > K



(39)
> >
= q > Q x̂> K + β q > Q ŷ > K





(40)

>
The first term q > Q x̂> K is unconditioned on β, and thus shared for every hβ . Let us focus



>
>
on the second term β q > Q ŷ > K . If q > Q ŷ > K > 0, then:

e (q, h2 ) > e (q, h1 ) (41)

21
Published as a conference paper at ICLR 2022

>
Otherwise, if q > Q ŷ > K


≤ 0:
e (q, h0 ) ≥ e (q, h1 ) (42)
Thus, for any query q, the key h1 can never get the highest score, and thus cannot be “selected”. That
is, the key h1 cannot satisfy that e (q, h1 ) is strictly greater than any other key.
In the general case, let h0 , h1 ∈ Rd be some non-zero vectors , and λ is some scalar such that
0 < λ < 1.
Consider the following linearly dependent set of vectors:
K = Q = {βh1 + (1 − β) h0 | β ∈ {0, λ, 1}} (43)

For any query q ∈ Q and β ∈ {0, λ, 1} we define:

p
s (q, β) = e (q, (βh1 + (1 − β) h0 )) · dk (44)
Where e is the attention score function of DPGAT (Equation (13)).
Therefore:

 >
>
s (q, β) = q > Q (βh1 + (1 − β) h0 ) K (45)

>
= q > Q βh> >


1 K + (1 − β) h0 K (46)

>
= q > Q βh> > >


1 K + h0 K − βh0 K (47)

>
= q > Q β h> > >



1 K − h0 K + h0 K (48)

>
>
= β q > Q h> >
+ q > Q h>



1 K − h0 K 0K (49)

>
If q > Q h> >


1 K − h0 K > 0:
e (q, h1 ) > e (q, hλ ) (50)

>
Otherwise, if q > Q h> >


1 K − h0 K ≤ 0:
e (q, h0 ) ≥ e (q, hλ ) (51)
Thus, for any query q, the key hλ cannot be selected. That is, the key hλ cannot satisfy that e (q, hλ )
is strictly greater than any other key. Therefore, there are mappings ϕ, for which dynamic attention
does not hold.
While we prove that GATv2 computes dynamic attention (Appendix A) for any set of node represen-
tations K = Q, there are sets of node representations and mappings ϕ for which dynamic attention
does not hold for DPGAT. Thus, DPGAT is strictly weaker than GATv2.

E.3 E MPIRICAL E VALUATION

Here we repeat the experiments of Section 4 with DPGAT. We remind that DPGAT is strictly weaker
than our proposed GATv2 (see a proof in Appendix E.1).

F S TATISTICAL S IGNIFICANCE
Here we report the statistical significance of the strongest GATv2 and GAT models of the experiments
reported in Section 4.

22
Published as a conference paper at ICLR 2022

72
DPGAT DPGAT
GATv2 GATv2
GAT 32 GAT
70
Accuracy

Accuracy
30
68

28
66

0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
p – noise ratio p – noise ratio
(a) ogbn-arxiv (b) ogbn-mag

Figure 8: Test accuracy compared to the noise ratio: GATv2 and DPGAT are more robust to structural
noise compared to GAT. Each point is an average of 10 runs, error bars show standard deviation.

Table 10: Accuracy (5 runs±stdev) on VAR M ISUSE. GATv2 is more accurate than all GNNs in both
test sets, using GAT’s hyperparameters. † – previously reported by Brockschmidt (2020).

Model SeenProj UnseenProj

†
No- GCN 87.2±1.5 81.4±2.3
Attention GIN† 87.1±0.1 81.1±0.9
GAT† 86.9±0.7 81.2±0.9
Attention DPGAT 88.0±0.8 81.5±1.2
GATv2 88.0±1.1 82.8±1.7

Table 11: Average accuracy (Table 11a) and ROC-AUC (Table 11b) in node-prediction datasets (10
runs±std). In all datasets, GATv2 outperforms GAT. † – previously reported by Hu et al. (2020).

(a) (b)

Model Attn. Heads ogbn-arxiv ogbn-products ogbn-mag ogbn-proteins

†
GCN 0 71.74±0.29 78.97±0.33 30.43±0.25 72.51±0.35
GraphSAGE† 0 71.49±0.27 78.70±0.36 31.53±0.15 77.68±0.20
1 71.59±0.38 79.04±1.54 32.20±1.46 70.77±5.79
GAT
8 71.54±0.30 77.23±2.37 31.75±1.60 78.63±1.62
1 71.52±0.17 76.49±0.78 32.77±0.80 63.47±2.79
DPGAT
8 71.48±0.26 73.53±0.47 27.74±9.97 72.88±0.59
1 71.78±0.18 80.63±0.70 32.61±0.44 77.23±3.32
GATv2 (this work)
8 71.87±0.25 78.46±2.45 32.52±0.39 79.52±0.55

23
Published as a conference paper at ICLR 2022

Table 12: Average error rates (lower is better), 5 runs ± standard deviation for each property, on the
QM9 dataset. The best result among GAT, GATv2 and DPGAT is marked in bold; the globally best
result among all GNNs is marked in bold and underline. † was previously tuned and reported by
Brockschmidt (2020).

Predicted Property
Model 1 2 3 4 5 6 7
†
GCN 3.21±0.06 4.22±0.45 1.45±0.01 1.62±0.04 2.42±0.14 16.38±0.49 17.40±3.56
GIN† 2.64±0.11 4.67±0.52 1.42±0.01 1.50±0.09 2.27±0.09 15.63±1.40 12.93±1.81
GAT1h 3.08±0.08 7.82±1.42 1.79±0.10 3.96±1.51 3.58±1.03 35.43±29.9 116.5±10.65
GAT8h † 2.68±0.06 4.65±0.44 1.48±0.03 1.53±0.07 2.31±0.06 52.39±42.58 14.87±2.88
DPGAT8h 2.63±0.09 4.37±0.13 1.44±0.07 1.40±0.03 2.10±0.07 32.59±34.77 11.66±1.00
DPGAT1h 3.20±0.17 8.35±0.78 1.71±0.03 2.17±0.14 2.88±0.12 25.21±2.86 65.79±39.84
GATv21h 3.04±0.06 6.38±0.62 1.68±0.04 2.18±0.61 2.82±0.25 20.56±0.70 77.13±37.93
GATv28h 2.65±0.05 4.28±0.27 1.41±0.04 1.47±0.03 2.29±0.15 16.37±0.97 14.03±1.39

Predicted Property Rel. to

Model 8 9 10 11 12 13 GAT8h
GCN† 7.82±0.80 8.24±1.25 9.05±1.21 7.00±1.51 3.93±0.48 1.02±0.05 -1.5%
GIN† 5.88±1.01 18.71±23.36 5.62±0.81 5.38±0.75 3.53±0.37 1.05±0.11 -2.3%
GAT1h 28.10±16.45 20.80±13.40 15.80±5.87 10.80±2.18 5.37±0.26 3.11±0.14 +134.1%
GAT8h † 7.61±0.46 6.86±0.53 7.64±0.92 6.54±0.36 4.11±0.27 1.48±0.87 +0%
DPGAT1h 12.93±1.70 13.32±2.39 14.42±1.95 13.83±2.55 6.37±0.28 3.28±1.16 +77.9%
DPGAT8h 6.95±0.32 7.09±0.59 7.30±0.66 6.52±0.61 3.76±0.21 1.18±0.33 -9.7%
GATv21h 10.19±0.63 22.56±17.46 15.04±4.58 22.94±17.34 5.23±0.36 2.46±0.65 +91.6%
GATv28h 6.07±0.77 6.28±0.83 6.60±0.79 5.97±0.94 3.57±0.36 1.59±0.96 -11.5%

72
GATv2 (p-value) GATv2 (p-value)
GAT GAT
0.

32
p-
00

va
02

0.

lu

<0

70
0

e<
<0

00

.0 <
0.

00 0.
Accuracy

Accuracy
.0

00
00 0.0

1 00
<0

p-

01 .000
1 00

va

<0
.0
<

lu
00 0.00

.0
e<

00

<0
1

<0

01

30
0. <0

.0
.0
<0 01 001

1
1

00 .0

68
00
00
.0

1
1
01

00
1

28
66

0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
noise ratio noise ratio
(a) ogbn-arxiv (b) ogbn-mag

Figure 9: Test accuracy and statistical significance compared to the noise ratio: GATv2 is more robust
to structural noise compared to GAT. Each point is an average of 10 runs, error bars show standard
deviation.
Table 13: Accuracy (5 runs±stdev) on VAR M ISUSE. GATv2 is more accurate than all GNNs in both
test sets, using GAT’s hyperparameters. † – previously reported by Brockschmidt (2020).

Model SeenProj UnseenProj

†
GAT 86.9±0.7 81.2±0.9
GATv2 88.0±1.1 82.8±1.7
p-value 0.048 0.049

24
Published as a conference paper at ICLR 2022

Table 14: Accuracy (100 runs±stdev) on Pubmed. GATv2 is more accurate than GAT.

Model Accuracy
GAT 78.1±0.59
GATv2 78.5±0.38
p-value < 0.0001

Table 15: Average accuracy (Table 15a) and ROC-AUC (Table 15b) in node-prediction datasets (30
runs±std). We report on the best GAT / GATv2 from Table 1.

(a) (b)

Model ogbn-arxiv ogbn-products ogbn-mag ogbn-proteins

GAT 71.65±0.38 79.04±1.54 32.36±1.10 78.29 ±1.59
GATv2 71.93±0.35 80.63±0.70 33.01±0.41 78.96±1.19
p-value 0.0022 <0.0001 0.0018 0.0349

Table 16: Average Hits@50 (Table 16a) and mean reciprocal rank (MRR) (Table 16b) in link-
prediction benchmarks from OGB (30 runs±std). We report on the best GAT / GATv2 from Table 3.

(a) (b)

ogbl-collab ogbl-citation2
Model w/o val edges w/ val edges
GAT 42.24±2.26 46.02±4.09 79.91±0.13
GATv2 43.82±2.24 49.06±2.50 80.20±0.62
p-value 0.0043 0.0005 0.0075

Table 17: Average error rates (lower is better), 20 runs ± standard deviation for each property, on the
QM9 dataset. We report on GAT and GATv2 with 8 attention heads.

Predicted Property
Model 1 2 3 4 5 6 7
GAT 2.74±0.08 4.73±0.40 1.47±0.06 1.53±0.06 2.44±0.60 55.21±42.33 25.36±31.42
GATv2 2.67±0.08 4.28±0.23 1.43±0.05 1.51±0.07 2.21±0.08 16.64±1.17 13.61±1.68
p-value 0.0043 <0.0001 0.0138 0.1691 0.0487 0.0001 0.0516

Predicted Property
Model 8 9 10 11 12 13
GAT 7.36±0.87 6.79±0.86 7.36±0.93 6.69±0.86 4.10±0.29 1.51±0.84
GATv2 6.13±0.59 6.33±0.82 6.37±0.86 5.95±0.62 3.66±0.29 1.09±0.85
p-value <0.0001 0.0458 0.0006 0.0017 <0.0001 0.0621

25
Published as a conference paper at ICLR 2022

G C OMPLEXITY A NALYSIS
We repeat the definitions of GAT, GATv2 and DPGAT:
e (hi , hj ) =LeakyReLU a> · [W hi kW hj ]


GAT (Veličković et al., 2018): (52)
>
GATv2 (our fixed version): e (hi , hj ) =a LeakyReLU (W · [hi khj ]) (53)

>  √
e (hi , hj ) = h> >


DPGAT (Vaswani et al., 2017): i Q · hj K / d0 (54)

G.1 T IME C OMPLEXITY

GAT As noted by Veličković et al. (2018), the time complexity of a single GAT head may be
expressed as O (|V|dd0 + |E|d0 ). Because of GAT’s static attention, this computation can be further
optimized, by merging the linear layer a1 with W , merging a2 with W , and only then compute
a>
{1,2} W hi for every i ∈ V.

GATv2 require the same computational cost as GAT’s declared complexity: O (|V|dd0 + |E|d0 ): we
0 0
denote W = [W1 kW2 ], where W1 ∈ Rd ×d and W2d ×d contain the left half and right half of the
columns of W , respectively. We can first compute W1 hi and W2 hj for every i, j ∈ V. This takes
O (|V|dd0 ).
Then, for every edge (j, i), we compute LeakyReLU (W · [hi khj ]) using the precomputed W1 hi
and W2 hj , since W · [hi khj ] = W1 hi + W2 hj . This takes O (|E|d0 ).
Finally, computing the results of the linear layer a takes additional O (|E|d0 ) time, and overall
O (|V|dd0 + |E|d0 ).
DPGAT also takes the same time. We can first compute h> >
i Q and hj K for every i, j ∈ V. This
>
takes O (|V|dd0 ). Computing the dot-product h> h>



i Q j K for every edge (j, i) takes additional
0 0 0
O (|E|d ) time, and overall O (|V|dd + |E|d ).

G.2 PARAMETRIC C OMPLEXITY

GAT GATv2 DPGAT

0 0 0 0
Official 2d + dd d + 2dd 2ddk + dd0
In our experiments 2d0 + dd0 d0 + dd0 2dd0

Table 18: Number of parameters for each GNN type, in a single layer and a single attention head.

All parametric costs are summarized in Table 18. All following calculations refer to a single layer
having a single attention head, omitting bias vectors.
0 0
GAT has learned vector and a matrix: a ∈ R2d and W ∈ Rd ×d , thus overall 2d0 + dd0 learned
parameters.
0
GATv2 has a matrix that is twice larger: W ∈ Rd ×2d , because it is applied on the concatenation
[hi khj ]. Thus, the overall number of learned parameters is d0 + 2dd0 . However in our experiments,
to rule out the increased number of parameters over GAT as the source of empirical difference, we
constrained W = [W 0 kW 0 ], and thus the number of parameters were d0 + dd0 .

DPGAT has Q and K matrices of sizes ddk each, and additional dd0 parameters in the value matrix
V , thus 2ddk + dd0 parameters overall. However in our experiments, we constrained Q = K and
set dk = d0 , and thus the number of parameters is only 2dd0 .

26