Graph-based Modeling of Online Communities for Fake News Detection

Shantanu Chandra♣ , Pushkar MishraF , Helen Yannakoudakis♠ , Madhav NimishakaviF ,

Marzieh SaeidiF , Ekaterina Shutova♣
♣
ILLC, University of Amsterdam, The Netherlands
F
Facebook AI, London, United Kingdom
♠
Dept. of Informatics, King’s College London, United Kingdom
[email protected], [email protected], [email protected],
{pushkarmishra, madhavn, marzieh }@fb.com

Abstract modeling the structure, style and content of a news

article (Khan et al., 2019; Pérez-Rosas et al., 2017),
Over the past few years, there has been a sub- no attempts have been made to understand and ex-
stantial effort towards automated detection of
arXiv:2008.06274v3 [cs.CL] 14 Sep 2020

ploit the online community that interacts with the

fake news on social media platforms. Exist-
ing research has modeled the structure, style, article.
content, and patterns in dissemination of on- To advance this line of research, we propose
line posts, as well as the demographic traits SAFER (Socially Aware Fake nEws detection
of users who interact with them. However, no fRamework), a graph-based approach to fake news
attention has been directed towards modeling detection that aggregates information from 1) the
the properties of online communities that in-
content of the article, 2) content-sharing behavior
teract with the posts. In this work, we pro-
pose a novel social context-aware fake news
of users who shared the article, and 3) the social
detection framework, SAFER, based on graph network of those users. We frame the task as a
neural networks (GNNs). The proposed frame- graph-based modeling problem over a heteroge-
work aggregates information with respect to: neous graph of users and the articles shared by
1) the nature of the content disseminated, 2) them. We perform a systematic comparison of sev-
content-sharing behavior of users, and 3) the eral graph neural network (GNN) models as graph
social network of those users. We furthermore encoders in our proposed framework and introduce
perform a systematic comparison of several
novel methods based on relational and hyperbolic
GNN models for this task and introduce novel
methods based on relational and hyperbolic GNNs, which have not been previously used for
GNNs, which have not been previously used user or community modeling within NLP. By using
for user or community modeling within NLP. relational GNNs, we explicitly model the different
We empirically demonstrate that our frame- relations that exist between the nodes of the het-
work yields significant improvements over ex- erogeneous graph, which the traditional GNNs are
isting text-based techniques and achieves state- not designed to capture. Furthermore, euclidean
of-the-art results on fake news datasets from
embeddings used by the traditional GNNs have a
two different domains.
high distortion when embedding real world hierar-
1 Introduction chical and scale-free graphs1 (Ravasz and Barabási,
2003; Chen et al., 2013). Thus, by using hyperbolic
The spread of fake news online leads to undesirable GNNs we capture the relative distance between the
consequences in many areas of societal life, notably node representations more precisely by operating
in the political arena and healthcare with the most in the hyperbolic space. Our methods generate rich
recent example being the COVID-19 “Infodemic” community-based representations for articles. We
(Zarocostas, 2020). Its consequences include politi- demonstrate that, when used alongside text-based
cal inefficacy, polarization of society and alienation representations of articles, SAFER leads to sig-
among individuals with high exposure to fake news nificant gains over existing methods for fake news
(Balmas, 2014; Norton and Greenwald, 2016). Re- detection and achieves state-of-the-art performance.
cent years have therefore seen a growing interest in
1
automated methods for fake news detection, which A Scale Free Network is one in which the distribution
is typically set up as a binary classification task. of links to nodes follows a power law, i.e., the vast majority
of nodes have very few connections, while a few important
While a large proportion of work has focused on nodes (hubs) have a huge number of connections.
We also make the code publicly available2 . tures derived from the article, news source, users
and their interactions and timeline of posting to de-
2 Related Work tect fake news. They construct two homogeneous
Approaches to fake news detection can be catego- sub-graphs (news-source and user sub-graph) and
rized into three different types: content-, propa- model them separately in an unsupervised setting
gation- and social-context based. Content-based for proximity relations. They also use the user’s
approaches model the content of articles, such stance in relation to the shared content as addi-
as the headline, body text, images and external tional information via a stance detection network
URLs. Some methods utilize knowledge graphs pre-trained on a self-curated dataset.
and subject-predicate-object triples (Ciampaglia Our formulation of the problem is distinct from
et al., 2015; Shi and Weninger, 2016), while other these methods in three ways. Firstly, we construct a
feature-based methods model writing style, psycho- single heterogeneous graph consisting of two kinds
linguistic properties of text, rhetorical relations and of nodes and edges and model them together in a
content readability (Popat, 2017; Castillo et al., semi-supervised graph learning setup. Secondly,
2011; Pérez-Rosas et al., 2017; Potthast et al., we do not perform user profiling, but rather com-
2017). Others use neural networks (Ma et al., pute community-wide social-context features, and
2016), with attention-based architectures such as to the best of our knowledge, no prior work has
HAN (Okano et al., 2020) and dEFEND (Shu et al., investigated the role of online communities in fake
2019a) outperforming other neural methods. Re- news detection. Third, to capture the role of com-
cent multi-modal approaches encoding both tex- munities, we only use the information about the
tual and visual features of news articles as well as users’ networks, without the need for any personal
tweets (Shu et al., 2019c; Wang et al., 2018), have information from user’s profile and yet outperform
advanced the performance further. the existing methods that incorporate those. Fur-
Propagation-based methods analyze patterns in thermore, since our methods do not use any user-
the spread of news based on news cascades (Zhou specific information, such as their location, race
and Zafarani, 2018) which are tree structures that or gender, they therefore do not learn to associate
capture the content’s post and re-post patterns. specific population groups with specific online be-
These methods make predictions in two ways: 1) haviour, unlike other methods that explicitly in-
computing the similarity between the cascades corporate user-specific features and their personal
(Kashima et al., 2003; Wu et al., 2015); or 2) rep- information by design. We believe the latter would
resenting news cascades in a latent space for clas- pose an ethical concern, which our techniques help
sification (Ma et al., 2018). However, they are not to alleviate.
well-suited to large social-network setting due to
3 Datasets
their computational complexity.
Social-context based methods employ the users’ For our experiments, we use fake news datasets
meta-information obtained from their social media from two different domains, i.e., celebrity gossip
profiles (e.g. geo-location, total words in profile de- and healthcare, to show that our proposed method
scription, etc.) as features for detecting fake news is domain-agnostic. All user information collected
(Shu et al., 2019b, 2020). Recently, several works for the experiments is de-identified.
have leveraged GNNs to learn user representations FakeNewsNet3 (Shu et al., 2018) is a a pub-
for other tasks, such as abuse (Mishra et al., 2019), licly available benchmark for fake news detection.
political perspective (Li and Goldwasser, 2019) and The dataset contains news articles from two fact-
stance detection (Del Tredici et al., 2019). checking sources, PolitiFact and GossipCop, along
Two works, contemporaneous to ours, have also with links to Twitter posts mentioning these arti-
proposed to use GNNs for the task of fake news cles. PolitiFact4 is a fact-checking website for po-
detection. Han et al. (2020) applied GNNs on a ho- litical statements; GossipCop5 is a website that fact-
mogeneous graph constructed in the form of news checks celebrity and entertainment stories. Gossip-
cascades by using just shallow user-level features Cop contains a substantially larger set of articles
such as no. of followers, status and tweet mentions.
3
On the other hand, Nguyen et al. (2020) use fea- https://tinyurl.com/uwadu5m
4
https://www.politifact.com/
2 5
https://github.com/shaanchandra/SAFER https://www.gossipcop.com/
 Training Testing
Graph Encoder

GNN
u1
Lgraph
u2
GNN u3
u4 sg
Ug
LR
Text
Text Text
LR
Classifier
Llr
Text 
ssafer
 Text Text

Test Sample Text Text  

RoBERTa
Text Encoder
st
 Text Text
 Text Text
Text  Text
 Text Text
TextText  Text
Text   RoBERTa
TextText  Text
Text  
Text Text  
Ltext Encode Aggregate Classify

Figure 1: Visual representation of the proposed SAFER framework. Graph and text encoders are trained independently followed
by training of a logistic regression (LR) classifier. During inference, the text of the article as well as information about its social
network of users are encoded by the trained text and graph encoders respectively. Finally, the social-context and textual features
of the article are concatenated for classification using the trained LR classifier.

compared to PolitiFact (over 21k news articles with 4 Methodology

text vs. around 900) and is therefore the one we use
in our experiments. We note that some articles have 4.1 Constructing the Community Graph
become unavailable over time. We also excluded
60 articles that are less than 25 tokens long. In For each of the datasets, we create a heteroge-
total, we work with 20, 350 articles of the original neous community graph G consisting of two sets
set (23% fake and 77% real). of nodes: user nodes Nu and article nodes Na . An
article node a ∈ Na is represented by a binary bag-
FakeHealth6 (Dai et al., 2020), is a publicly avail- of-words (BOW) vector a = [w1 , .., wj , .., w|V | ],
able benchmark for fake news detection specifically where |V| is the vocabulary size and wi ∈ {0, 1}.
in the healthcare domain. The dataset is collected A user node u ∈ Nu is represented by a binary
from healthcare information review website Health BOW vector constructed over all the articles that
News Review7 , which reviews whether a news arti- they have shared: u = [a1 | a2 | ... | aM ], where |
cle is reliable according to 10 criteria and gives it denotes the element-wise logical OR and M is the
a score from 1-5. In line with the original authors total number of articles shared by the user. Next,
of the dataset, we consider an article as fake for we add undirected edges of two types: 1) between a
scores less than 3 and real otherwise. The dataset is user and an article node if the user shared the article
divided into two datasets based on the nature of the in a tweet/retweet (article nodes may therefore be
source of the articles. HealthStory contains articles connected to multiple user nodes), and 2) between
that are news stories, i.e., reported by news media two user nodes if there is a follower–following rela-
such as Reuters Health. HealthRelease contains tionship between them on Twitter.8 We work with
articles that are news releases from various insti- the “top N most active users” (N=20K for Health-
tutions such as universities, research centers and Story, N=30K for GossipCop) subset and motivate
companies. HealthStory contains a considerably this decision in Section 6.2. To avoid effects of any
larger set of articles compared to HealthRelease bias from frequent users, we exclude users who
(over 1600 vs around 600) and is therefore the one have shared more than 30% of the articles in ei-
we use in our experiments. We again note that ther class. The resulting graph has 29, 962 user
some articles have become unavailable over time. nodes, 16, 766 article nodes (articles from test set
We also exclude 27 articles that are less than 25 excluded) and over 1.2M edges for GossipCop.
tokens long. In total, we work with 1611 of the Meanwhile, the HealthStory community graph con-
original set (28% fake and 72% real articles). tains 12, 266 user nodes, 1291 article nodes (test
set articles excluded) and over 450K edges.
6
https://tinyurl.com/y36h42zu
7 8
https://www.healthnewsreview.org/ We use Twitter APIs to retrieve the required information.
4.2 SAFER: Fake news detection framework Convolutional neural network (CNN). We adopt
the sentence-level encoder of Kim (2014) at the
The proposed framework – detailed below and vi-
document level. This model uses multiple 1-D
sualized in Figure 1 – employs two components
convolution filters of different sizes that aggregate
in its architecture, namely graph- and text-based
information by sliding over the length of the article.
encoders, and its working can be broken down into
The final fixed-length article representation is ob-
two phases: training and testing.
tained via max-over-time pooling over the feature
Training phase: We first train the graph and text maps.
encoders independently on the training set. The RoBERTa. As our main text encoder, we fine-tune
input to the text encoder is the text of the article the transformer encoder architecture, RoBERTa
and it is trained on the task of article classification (Liu et al., 2019b), and use it for article classifica-
for fake news detection. The trained text encoder tion. RoBERTa is a language model pre-trained
generates the text-based features of the article con- with dynamic masking. Specifically, we use it to
tent st ∈ Rdt where dt is the hidden dimension of encode the first 512 tokens of each article and use
the text encoder. The graph encoder is a GNN that the [ C L S ] token as the article embedding for clas-
takes as input the community graph (constructed sification.
as detailed in §4.1). The GNN is trained with su-
pervised loss from the article nodes that is back- 4.4 Graph Encoders
propagated to the rest of the network. The trained
We experiment with six different GNN architec-
GNN is able to generate a set of user embeddings
tures for generating user embeddings as detailed
Ug = {u1 , u2 , ..., um } where ui ∈ Rdg , dg is the
below:
hidden dimension of the graph encoder and m is the
total number of users that interacted with the article. Graph Convolution Networks (GCNs). GCNs
These individual user representations are then ag- (Kipf and Welling, 2016) take as input a graph G
gregated into a single fixedP size vector defined by its adjacency matrix A ∈ Rn×n (where

via normal-
m dg n is number of nodesP in the graph)9 , a degree matrix
ized sum such that sg = i=1 i /m, sg ∈ R
u
where sg denotes the social-context features of the D such that Dii = j Aij , and a feature matrix
article. The final social context-aware representa- F ∈ Rn×m containing the m-dimensional feature
tion of the article is computed as ssaf er = sg ⊕ st , vectors for the nodes. The recursive propagation
where ⊕ is the concatenation operator. This form of step of a GCN at the ith convolutional layer is given
aggregation helps SAFER to retain the information by: Oi = σ(ÃO(i−1) W i ) where σ denotes an ac-
1 1
that each representation encodes about different tivation function, Ã = D− 2 AD− 2 is the degree-
aspects of the shared content. Finally, ssaf er is normalized adjacency matrix; W i ∈ Rti−1 ×ti is
used to train a logistic regression (LR) classifier the weight matrix of the ith convolutional layer;
on the training set. Intuitively, the trained text en- O(i−1) ∈ Rn×ti−1 represents the output of the pre-
coder captures the linguistic cues from the content ceding convolution layer and ti is the number of
that are crucial for the task. Similarly, the trained hidden units in the ith layer, with t0 = m.
graph encoder learns to assign users to implicit on- Graph Attention Networks (GAT). GAT
line communities based on their content-sharing (Veličković et al., 2017) is a non-spectral architec-
patterns and social connections. ture that leverages the spatial information of a node
Testing phase: To classify unseen content as fake directly by learning different weights for different
or real, SAFER takes as input the text of the article nodes in a neighborhood using a self-attention
as well as the network of users that interacted with mechanism. GAT is composed of graph attention
it. It then follows the same procedure as detailed layers. In each layer, a shared, learnable linear
above to generate the social context-aware repre- transformation W ∈ Rti−1 ×ti is applied to the
sentation of the to-be-verified test article, ssaf er , input features of every node, where ti is the number
and uses the trained LR classifier to classify it. of hidden units in layer i. Next, self-attention
is applied on nodes, where a shared attention
4.3 Text Encoders mechanism computes attention coefficients euv
between pairs of nodes to indicate the importance
We experiment with two different architectures as
9
text encoders in SAFER: A is symmetric, i.e., Aij = Aji , with self-loops Aii = 1.
of the features of node v to node u. To inject graph vectors of first-order neighbor nodes through a nor-
structural information, masked attention is applied malized sum. R-GAT also follows the same setup,
by computing euv only for nodes v ∈ U(u) that except the aggregation is done using the graph at-
are in the first-order neighborhood of node u. The tention layer as described in GAT. This architecture
final node representation is obtained by linearly helps us to aggregate information from user and ar-
combining normalized attention coefficients with ticle nodes selectively from our community graph.
their corresponding neighborhood node features. Hyperbolic GCN / GAT. Chami et al. 2019 build
GraphSAGE. SAGE (Hamilton et al., 2017) is an upon previous work (Liu et al., 2019a; Ganea et al.,
inductive framework that learns aggregator func- 2018) to combine the expressiveness of GCN/GAT
tions that generate node embeddings from a node’s with hyperbolic geometry to learn improved repre-
local neighborhood. First, each node u ∈ G ag- sentations for scale-free graphs. Hy-GCN/GAT
gregates information (through either mean, sum or first map the euclidean input to the hyperbolic
pooling) from its local neighborhood hk−1
v , ∀v ∈ space (we use the Poincar ball model), which is
k−1
U(u) into a single vector hU (u) where k denotes the Riemannian manifold with constant negative
the depth of the search, hk denotes the node’s rep- sectional curvature -1/K. Next, analogous to the
resentation at that step and U(u) is set of neighbor mean aggregation by the GCN, Hy-GCN computes
nodes of u. Next, it concatenates the node’s cur- the Frchet mean (Fréchet, 1948) of a node’s neigh-
rent representation hk−1 with that of its aggregated bours’ embeddings while the Hy-GAT performs
u
k−1
neighborhood vector hU (u) . This vector is then aggregation in tangent spaces using hyperbolic at-
passed through a multi-layer perceptron (MLP) tention. Finally, Hy-GCN/GAT use hyperbolic non-
Ki−1 ,Ki
with non-linearity to obtain the new node repre- linear activation function σ ⊗ given the hy-
sentation hku to be used at depth k + 1. Once the perbolic curvatures -1/Ki−1 , -1/Ki at layers i − 1
aggregator weights are learned, the embedding of and i where ⊗ is the Mbius scalar multiplication
an unseen node can be generated from its features operator. This is crucial as it allows the model to
and neighborhood. smoothly vary curvature at each layer.
Relational GCN/GAT. R-GCN (Schlichtkrull
4.5 Baselines and Comparison Systems
et al., 2018) and R-GAT are an extension of GCN
and GAT for relational data and build upon the tra- We compare the performance of the proposed
ditional differentiable message passing framework. framework with seven supervised classification
The networks accept input in the form of a graph methods: two purely text-based baselines, a user-
G = (V, E, R) where V denotes the set of nodes, sharing majority voting baseline, a GNN-based “so-
E denotes the set of edges connecting the nodes cial baseline” and three architectures from the liter-
and R denotes the edge relations (u, r, v) ∈ E ature.
where r ∈ R is a relation type and u, v ∈ V . The Baselines. The setup for the baselines is detailed
R-GCN forward pass update step is: below:
1. Text-baselines. We use the CNN and
   RoBERTa architectures described earlier to obtain
X X 1
h(i)
u = σ
 (i−1) 
Wr(i−1) hl
(i−1) (i−1) 
+ W0 hu article representations. The input to the CNN en-
cu,r
r
r∈R l∈Uu coder is ELMo embeddings (Peters et al., 2018) of
the article tokens, while RoBERTa uses its own
(i)
where hu is the final node representation of node tokenizer to generate initial token representations.
r
u at layer i, Uu denotes the set of neighbor indices 2. Majority sharing baseline. This simple base-
of node u under relation r ∈ R, Wr is the relation- line classifies articles as fake or real based on the
specific trainable weight parameter and cu,r is a sharing statistics of users that tweeted or retweeted
task specific normalization constant that can either about it. If, on average, the users that interact with
be learned or set in advance (such as cu,r = |Uur |). an article have shared more fake articles, then the
Note that each node’s feature at layer i is also in- article is tagged as fake, and real otherwise.
formed of its features from layer i − 1 by adding 3. Social Baseline. We introduce a graph-based
a self-loop to the data with a relation type learned model that measures the effectiveness of purely
using the trainable parameter W0 . Intuitively, this structural aspects of the community graph captured
propagation step aggregates transformed feature by the GNNs (without access to text). The user
node embeddings are constructed as described ear- using the AdamW (Loshchilov and Hutter, 2017)
lier, but with the article nodes being initialized ran- optimizer (except for Hy-GCN/-GAT that use Rie-
domly. Here, the community-based features solely mannian Adam; Bécigneul and Ganea 2018) with
capture properties of the network. The classifica- an early stopping patience of 10. For GossipCop,
tion is done using just the social-context feature by we use a learning rate of 5 · 10−3 for Hy-GCN/-
an LR classifier. GAT; 1 · 10−4 for SAGE and R-GAT; 1 · 10−3 for
Comparison Systems. We compare the perfor- R-GCN; and 5 · 10−4 for the rest. We use weight
mance of the proposed framework with three meth- decay of 5 · 10−1 for RoBERTA; 2 · 10−3 for SAGE
ods from literature: and R-GCN; 1 · 10−3 for the rest. We use dropout
1. HAN (Shu et al., 2019a). Hierarchical atten- of 0.4 for GAT and R-GCN; 0.2 for SAGE and
tion network first generates sentence embeddings R-GAT; 0.5 for CNN; and 0.1 for the rest. We use
using attention over (GRU-based) contextualised node masking probability of 0.1 for all the GNNs
word vectors. An article embedding is then ob- and attention dropout of 0.4 for RoBERTa. Finally,
tained in a similar manner by passing sentence we use a hidden dimension of 128 for SAGE; 256
vectors through a GRU and applying attention over for GCN and Hy-GCN; and 512 for the rest. Mean-
the hidden states. while for HealthStory, we use a learning rate of
2. dEFEND (Shu et al., 2019a). This method 1 · 10−4 for SAGE; 1 · 10−3 for R-GAT; 5 · 10−3
exploits contents of articles alongside comments GCN, Hy-GCN/GAT; and 5 · 10−4 for the rest. We
from users. Comment embeddings are obtained use weight decay of 5 · 10−1 for RoBERTa; 2 · 10−3
from a single layer bi-GRU and article embeddings for GAT, SAGE and R-GCN; and 1 · 10−3 for the
are generated using HAN. A cross-attention mecha- rest. We use dropout of 0.4 for GCN; 0.1 for Hy-
nism is applied over the two embeddings to exploit GCN/GAT and RoBERTa, 0.5 for CNN; and 0.2 for
users’ opinions and stance to better detect fake the rest. We use node masking probability of 0.2
news. for GAT and R-GCN; 0.3 for Hy-GCN/-GAT; and
3. SAFE (Zhou et al., 2020): This method uses 0.1 for the rest. Finally, we use attention dropout of
visual and textual features of the content. It uses 0.4 for RoBERTa and a hidden dimension of 128
a CNN to encode the textual as well as visual con- for SAGE; 256 for Hy-GAT and 512 for the rest.
tent of an article by initially processing the visual Results. The mean F1 scores for all models are
information using a pre-trained image2sentence summarized in Table 1. We note that the simple
model10 . It then concatenates these representations majority sharing baseline achieves an F1 of 77.19
to better detect fake news. on GossipCop while just 8.20 on HealthStory. This
highlights the difference in the content sharing be-
5 Experiments and Results
havior of users between the two datasets and we
Experimental setup. We use 70%, 10% and 20% explore this further in Section 6.3. We can also
of the total articles as train, validation and test splits see this difference in the strength of social context
respectively for both datasets. For CNN we use 128 information between the 2 datasets from the per-
filters of sizes [3,4,5] each. For HAN and dEFEND formance of the social baseline. Social baseline
we report the results in Shu et al. (2019a), while variants of all GNNs significantly (p < 0.05 under
for SAFE in Zhou et al. (2020). We use the large paired t-test) outperform all text-based methods in
version of RoBERTa and fine-tune all layers. Due to case of GossipCop but not in case of HealthStory.
class imbalance, we weight the loss from the fake However, all the social baselines outperform the
class 3 times more (in line with the class frequency majority sharing baseline demonstrating the contri-
in each of the datasets) while optimizing the binary bution of GNNs beyond capturing just the average
cross entropy loss of the sigmoid output from a 2- sharing behavior of interacting users. Note that we
layer MLP in all our experiments. We use dropout observe similar trends in experiments with CNN as
(Srivastava et al., 2014), attention dropout and node the text-encoder of the proposed framework.
masking (Mishra et al., 2020) for regularization. Finally, in case of GossipCop, all the variants
We use 2-layer deep architectures for all the GNNs. of the proposed SAFER framework significantly
For Hy-GCN/-GAT we train with learnable curva- outperform all their social baseline counterparts as
ture. We run all experiments with 5 random seeds well as all the text-based models. The relational
10
https://tinyurl.com/y3s965o5 GNN variants significantly outperform all the meth-
 Model GossipCop HealthStory
†
HAN 67.20 -
dEFEND† 75.00 -
Text SAFE
‡
89.50 -
CNN 66.73 53.81
R o BERT a 68.55 57.54
Maj. sharing baseline 77.19 8.20
Social baseline
SAGE 87.11 43.05
GCN 88.37 44.86
GAT 87.94 46.13
R - GCN 89.68 46.28
R - GAT 89.21 46.89
H y- GCN 87.45 44.90
H y- GAT 85.56 43.09
Graph
SAFER
SAGE 93.32 58.34 Figure 2: t-SNE plots of test article embeddings produced by
GCN 93.61 58.65
GAT 93.65 58.55 RoBERTa alone (top) and SAFER (R-GCN) (bottom). Fake
R - GCN 94.69 61.71 articles are in red and real in black
R - GAT 94.53 62.54
H y- GCN 93.64 61.81
H y- GAT 92.97 61.91
and real classes when the community features are
Table 1: F1 scores (fake class) on GossipCop and Health- combined with the textual features of the articles
Story. († ) denotes results reported from Shu et al. (2019a) compared to using textual features alone.
and (‡ ) from Zhou et al. (2020). Bold figure denotes signifi-
cantly better than other methods for that dataset. Underscore 6 Analysis
figures denote significantly better scores than baselines but
not significantly different from each other. 6.1 Effects of graph sparsity and frequent
users
Graph sparsity can affect the performance of GNNs
ods, while the hyperbolic variants perform on par as they rely on node connections to share informa-
with the traditional GNNs. In case of HealthStory, tion during training. Additionally, the presence of
we see that the traditional GNN variants signifi- frequent users that share many articles of a particu-
cantly outperform their social baseline counterparts lar class may introduce a bias in the model. In such
but not the best-performing text-based baseline (i.e., cases, the network may learn to simply map a user
RoBERTa). However, the relational and hyperbolic to a class and use that as a shortcut for classifica-
GNNs significantly outperform all other methods. tion. To investigate the effects of these phenomena,
Overall, we see that the proposed relational we perform an ablation experiment on GossipCop
GNNs outperform the traditional GNN models, in- by removing the more frequent/active users from
dicating the importance of modelling the different the graph in a step-wise fashion. This makes the
relations between nodes of a heterogeneous graph graph more sparse and discards many connections
separately. Hyperbolic GNNs are more expressive that the network could have learned to overfit on.
in embedding graphs that have a (deep) hierarchical Table 2 shows the performance of the GNN mod-
structure. Due to the nature of the datasets and limi- els when users sharing more than 10%, 5% and
tation of Twitter API (all retweets are mapped to the 1% articles of each class are removed from the
same source tweet, rather than forming a tree struc- graph. We see that the performance drops as users
ture), the community graph is just 2 levels deep. are removed successively; however, SAFER still
Thus, hyperbolic GNNs perform similar to the tra- outperforms all the text-based methods under the
ditional GNNs under our 2-layer setup. However, 10% and 5% setting, even without the presence of
if more social information were available, resulting a possible bias introduced by frequent users. For
in a deeper graph, we expect Hy-GNNs to exhibit the 1% setting, only the hyperbolic GNNs outper-
a superior performance. form the baselines and this setting illustrates that
In Figure 2, we use t-SNE (Maaten and Hinton, under extremely sparse conditions (65% of original
2008) to visualize the test articles representations density of an already sparse graph), the R-GNNs
generated by RoBERTa and SAFER (R-GCN). We struggle to learn informative user representations.
see a much cleaner and compact segregation of fake Overall, we see that Hy-GNNs are resilient to user
 Optimum threshold for top N users
Setting ρ SAFER F1 62.5
val-F1
R - GCN 81.87 60.0 test-F1
R - GAT 82.27
excl.>10% 0.78 H y- GCN 82.16
57.5

F1(fake) per split

H y- GAT 81.81 55.0

52.5
R - GCN 77.16
R - GAT 77.32 50.0
excl.>5% 0.71
H y- GCN 77.13 47.5
H y- GAT 77.01
45.0
R - GCN 65.89
All Top60k Top40k Top20k Top8k Top6k
R - GAT 65.32 Top N user subsets
excl.>1% 0.65 H y- GCN 71.99
H y- GAT 72.05 Figure 3: Validation and test set performance of the SAFER
(GCN) framework over varying subsets of most active users
Table 2: Results of SAFER variants on varying subsets of on HealthStory.
user nodes on GossipCop. ρ denotes relative graph density.

Bottleneck is the phenomenon of “over-

biases (if any) and can perform well even on sparse squashing of information from exponentially many
graphs. neighbours into small fixed-size vectors (Alon and
Yahav, 2020). Since each article is shared by many
6.2 Optimum support for effective learning
users and each user is connected to many other
GNNs learn node representations by aggregating users, the network can suffer from bottleneck which
information from their local neighborhoods. If the affects learning. In a 2-layer GNN setup, the ef-
unsupervised nodes have very sparse connections fective aggregation neighborhoods of each article
(e.g., users that have shared just one (or very few) node exponentially increases, as it aggregates in-
article(s)), then there is not enough support to learn formation from all the nodes that are within 2-hops
their social-context features from. The effective away from it.
neighborhood that a node uses is determined by the Due to these observations during our initial ex-
number of successive iterations of message passing periments, we choose to use just the “top N most
steps, i.e., the number of layers in a GNN. Thus, in active users”. We define “active users” as those that
principle we can add more GNN layers to enable have shared more articles, i.e., have sufficient sup-
the sparse unsupervised nodes to gain access to port to learn from, and hence can help us capture
back-propagating losses from distant supervised their content-sharing behavior better. In Figure 3
nodes too. However, simply stacking more layers we show the validation and test performance over
leads to various known problems of training deep varying subsets of active users in HealthStory. We
neural networks in general (vanishing gradients and see that as we successively drop the least active
overfitting due to large no. of parameters), as well users, the validation and test scores show a posi-
as graph specific problems such as over-smoothing tive trend. This illustrates the effects of bottleneck
and bottleneck phenomenon. on the network. However, the scores drop after a
Over-smoothing is the phenomenon where node certain threshold of users. This threshold is the
features tend to converge to the same vector and optimum number of users required to learn effec-
become nearly indistinguishable as the result of tively using the GNNs – adding more users leads
applying multiple GNN layers (Oono and Suzuki, to bottleneck and removing users leads to underfit-
2019; NT and Maehara, 2019). Moreover, in so- ting due to the lack of sufficient support to learn
cial network graphs, predictions typically rely only from. We see that validation and test scores are
on short-range information from the local neigh- correlated in this behavior and we tune our opti-
bourhood of a node and do not improve by adding mal threshold of users for effective learning based
distant information. In our community graph, mod- on the validation set of SAFER (GCN) and use
eling information from 2 hops away is sufficient to the same subset for all the other GNN encoders
aggregate useful community-wide information at for fair comparison. The best validation score was
each node and can be achieved with 2-layer GNNs. achieved at the top20K subset of most active users
Thus, learning node representations for sparsely while the test scores peaked for the top8K setting.
connected nodes from these shallow GNNs is chal- Thus, we run all our experiments with the top 20K
lenging. active users for HealthStory, and similarly top30K
 likely to share articles of one class predominantly.
This again restricts the GNNs from learning infor-
mative representations of these users, as it struggles
to assign them to any specific community due to
mixed signals.

7 Conclusion
We presented a graph-based approach to fake news
Figure 4: Article sharing behavior of 3 kinds of users (left) detection which leverages information-spreading
and; Average of real and fake article shares of type (c) users behaviour of social media users. Our results demon-
(right). strate that incorporating community-based model-
ing leads to substantially improved performance in
this task as compared to purely text-based models.
for GossipCop.
The proposed relational GNNs for user/community
6.3 Effect of article sharing patterns modeling outperformed the traditional GNNs indi-
cating the importance of explicitly modeling the
As discussed earlier, results in Table 1 show that
relations in a heterogeneous graph. Meanwhile,
there is a difference in the article sharing behavior
the proposed hyperbolic GNNs performed on par
of users between the two datasets. To understand
with other GNNs and we leave their application for
user characteristics better, we visualize the article
user/community modeling to truly hierarchical so-
sharing behavior of users for both the datasets in
cial network datasets as future work. In the future,
Figure 4. We visualize the composition of 3 types
it would be interesting to apply these techniques to
of users in the datasets: (a) users that share only
other tasks, such as rumour detection and modeling
real articles, (b) only fake articles and (c) those that
changes in public beliefs.
share articles from both classes. We see that the
majority of the users are type (c) in both datasets
(57.18% for GossipCop and 74.15% for Health-
Story). However, 38% of the users are type (b) in
GossipCop while just 9.96% in HealthStory. Fur-
thermore, we visualize the average of real and fake
articles shared by type (c) users on the right in Fig-
ure 4. From these observations, we note that the
GNNs are better positioned to learn user represen-
tations to detect fake articles in case of GossipCop,
since: (1) The community graph has enough sup-
port of type (b) users (38%). This aids the GNNs
to learn rich community-level features of users that
aid in detecting fake articles; (2) Even of the 57%
type (c) users, they are much more likely to share ar-
ticles of a single class (here, real). This again helps
the network to learn distinct features for these users
and assign them to a specific community.
However, in case of HealthStory, the GNNs
struggle to learn equally rich user representations
to detect fake articles since: (1) The community
graph has only around 10% of type (b) users. This
limits the GNNs from learning expressive commu-
nity level features for users that are more likely to
share fake articles and thereby are not able to use
them for accurate prediction. (2) A vast majority of
users (74%) share articles of both classes. To add
to that, these bulk of users are considerably less
References Yi Han, Shanika Karunasekera, and Christopher
Leckie. 2020. Graph neural networks with continual
Uri Alon and Eran Yahav. 2020. On the bottleneck of learning for fake news detection from social media.
graph neural networks and its practical implications.
arXiv preprint arXiv:2006.05205. George Karypis and Vipin Kumar. 1998. A fast and
high quality multilevel scheme for partitioning irreg-
Meital Balmas. 2014. When fake news becomes real: ular graphs. SIAM Journal on scientific Computing,
Combined exposure to multiple news sources and 20(1):359–392.
political attitudes of inefficacy, alienation, and cyn-
icism. Communication research, 41(3):430–454.
Hisashi Kashima, Koji Tsuda, and Akihiro Inokuchi.
Gary Bécigneul and Octavian-Eugen Ganea. 2018. Rie- 2003. Marginalized kernels between labeled graphs.
mannian adaptive optimization methods. arXiv In Proceedings of the 20th ICML (ICML-03), pages
preprint arXiv:1810.00760. 321–328.

Carlos Castillo, Marcelo Mendoza, and Barbara Junaed Younus Khan, Md Khondaker, Tawkat Islam,
Poblete. 2011. Information credibility on twitter. In Anindya Iqbal, and Sadia Afroz. 2019. A bench-
Proceedings of the 20th international conference on mark study on machine learning methods for fake
World wide web, pages 675–684. news detection. arXiv preprint arXiv:1905.04749.

Ines Chami, Rex Ying, Christopher R, and Jure Yoon Kim. 2014. Convolutional neural networks
Leskovec. 2019. Hyperbolic graph convolutional for sentence classification. In Proceedings of the
neural networks. 2014 Conference on Empirical Methods in Natural
Language Processing (EMNLP), pages 1746–1751.
Wei Chen, Wenjie Fang, Guangda Hu, and Michael W ACL.
Mahoney. 2013. On the hyperbolicity of small-
world and treelike random graphs. Internet Mathe- Thomas N Kipf and Max Welling. 2016. Semi-
matics, 9(4):434–491. supervised classification with graph convolutional
networks. arXiv preprint arXiv:1609.02907.
Wei-Lin Chiang, Xuanqing Liu, Si Si, Yang Li, Samy
Bengio, and Cho-Jui Hsieh. 2019. Cluster-gcn: An
efficient algorithm for training deep and large graph Chang Li and Dan Goldwasser. 2019. Encoding social
convolutional networks. In Proceedings of the 25th information with graph convolutional networks for-
ACM SIGKDD International Conference on Knowl- Political perspective detection in news media. In
edge Discovery & Data Mining, pages 257–266. Proceedings of the 57th Annual Meeting of the ACL,
pages 2594–2604. ACL.
Giovanni Luca Ciampaglia, Prashant Shiralkar, Luis M
Rocha, Johan Bollen, Filippo Menczer, and Alessan- Qi Liu, Maximilian Nickel, and Douwe Kiela. 2019a.
dro Flammini. 2015. Computational fact checking Hyperbolic graph neural networks. In H. Wal-
from knowledge networks. PloS one, 10(6). lach, H. Larochelle, A. Beygelzimer, F. dÁlché-Buc,
E. Fox, and R. Garnett, editors, Advances in Neu-
Enyan Dai, Yiwei Sun, and Suhang Wang. 2020. Gin- ral Information Processing Systems 32, pages 8230–
ger cannot cure cancer: Battling fake health news 8241. Curran Associates, Inc.
with a comprehensive data repository. In Proceed-
ings of the International AAAI Conference on Web Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man-
and Social Media, volume 14, pages 853–862. dar Joshi, Danqi Chen, Omer Levy, Mike Lewis,
Luke Zettlemoyer, and Veselin Stoyanov. 2019b.
Marco Del Tredici, Diego Marcheggiani, Sabine Roberta: A robustly optimized bert pretraining ap-
Schulte im Walde, and Raquel Fernández. 2019. proach. arXiv preprint arXiv:1907.11692.
You shall know a user by the company it keeps: Dy-
namic representations for social media users in NLP. Ilya Loshchilov and Frank Hutter. 2017. Decoupled
In Proceedings of the 2019 Conference on Empirical weight decay regularization.
Methods in Natural Language Processing and the
9th International Joint Conference on Natural Lan-
guage Processing (EMNLP-IJCNLP), pages 4707– Jing Ma, Wei Gao, Prasenjit Mitra, Sejeong Kwon,
4717. ACL. Bernard J. Jansen, Kam-Fai Wong, and Meeyoung
Cha. 2016. Detecting rumors from microblogs with
Maurice Fréchet. 1948. Les éléments aléatoires de na- recurrent neural networks. In Proceedings of the
ture quelconque dans un espace distancié. In An- Twenty-Fifth International Joint Conference on Arti-
nales de l’institut Henri Poincaré, volume 10, pages ficial Intelligence, IJCAI16, page 38183824. AAAI
215–310. Press.

Octavian Ganea, Gary Bécigneul, and Thomas Hof- Jing Ma, Wei Gao, and Kam-Fai Wong. 2018. Ru-
mann. 2018. Hyperbolic neural networks. In Ad- mor detection on twitter with tree-structured recur-
vances in neural information processing systems, sive neural networks. In Proceedings of the 56th An-
pages 5345–5355. nual Meeting of the ACL (Volume 1: Long Papers),
pages 1980–1989. ACL.
Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017.
Inductive representation learning on large graphs. In Laurens van der Maaten and Geoffrey Hinton. 2008.
Advances in neural information processing systems, Visualizing data using t-sne. Journal of machine
pages 1024–1034. learning research, 9(Nov):2579–2605.
Pushkar Mishra, Marco Del Tredici, Helen Yan- Baoxu Shi and Tim Weninger. 2016. Discriminative
nakoudakis, and Ekaterina Shutova. 2019. Abusive predicate path mining for fact checking in knowl-
Language Detection with Graph Convolutional Net- edge graphs. Knowledge-based systems, 104:123–
works. In Proceedings of the 2019 Conference of the 133.
North American Chapter of the ACL: Human Lan-
guage Technologies, Volume 1 (Long and Short Pa- Kai Shu, Limeng Cui, Suhang Wang, Dongwon Lee,
pers), pages 2145–2150. ACL. and Huan Liu. 2019a. Defend: Explainable fake
news detection. In Proceedings of the 25th ACM
Pushkar Mishra, Aleksandra Piktus, Gerard Goossen, SIGKDD International Conference on Knowledge
and Fabrizio Silvestri. 2020. Node masking: Mak- Discovery & Data Mining, KDD 19, page 395405.
ing graph neural networks generalize and scale bet- Association for Computing Machinery.
ter. ArXiv, abs/2001.07524.
Kai Shu, Deepak Mahudeswaran, Suhang Wang, Dong-
Van-Hoang Nguyen, Kazunari Sugiyama, Preslav won Lee, and Huan Liu. 2018. Fakenewsnet: A data
Nakov, and Min-Yen Kan. 2020. Fang: Leveraging repository with news content, social context and dy-
social context for fake news detection using graph namic information for studying fake news on social
representation. media. arXiv preprint arXiv:1809.01286.

Ben Norton and Glenn Greenwald. 2016. Washington Kai Shu, Suhang Wang, and Huan Liu. 2019b. Beyond
Post disgracefully promotes a McCarthyite Blacklist news contents: The role of social context for fake
from a hidden, new and very shady group. news detection. In Proceedings of the Twelfth ACM
International Conference on Web Search and Data
Hoang NT and Takanori Maehara. 2019. Revisiting Mining, pages 312–320.
graph neural networks: All we have is low-pass fil-
ters. arXiv preprint arXiv:1905.09550. Kai Shu, Guoqing Zheng, Yichuan Li, Subhabrata
Mukherjee, Ahmed Hassan Awadallah, Scott Rus-
Emerson Yoshiaki Okano, Zebin Liu, Donghong Ji, and ton, and Huan Liu. 2020. Leveraging multi-source
Evandro Eduardo Seron Ruiz. 2020. Fake news de- weak social supervision for early detection of fake
tection on fake.br using hierarchical attention net- news. arXiv preprint arXiv:2004.01732.
works. In Computational Processing of the Por-
tuguese Language, pages 143–152. Springer Inter- Kai Shu, Xinyi Zhou, Suhang Wang, Reza Zafarani,
national Publishing. and Huan Liu. 2019c. The role of user profiles for
fake news detection. In Proceedings of the 2019
IEEE/ACM International Conference on Advances
Kenta Oono and Taiji Suzuki. 2019. Graph neural net- in Social Networks Analysis and Mining, pages 436–
works exponentially lose expressive power for node 439.
classification. arXiv preprint arXiv:1905.10947.
Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky,
Verónica Pérez-Rosas, Bennett Kleinberg, Alexan- Ilya Sutskever, and Ruslan Salakhutdinov. 2014.
dra Lefevre, and Rada Mihalcea. 2017. Auto- Dropout: a simple way to prevent neural networks
matic detection of fake news. arXiv preprint from overfitting. The journal of machine learning
arXiv:1708.07104. research, 15(1):1929–1958.
Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Petar Veličković, Guillem Cucurull, Arantxa Casanova,
Gardner, Christopher Clark, Kenton Lee, and Luke Adriana Romero, Pietro Lio, and Yoshua Bengio.
Zettlemoyer. 2018. Deep contextualized word repre- 2017. Graph attention networks. arXiv preprint
sentations. In Proceedings of the 2018 Conference arXiv:1710.10903.
of the North American Chapter of the ACL: Human
Language Technologies, Volume 1 (Long Papers), Yaqing Wang, Fenglong Ma, Zhiwei Jin, Ye Yuan,
pages 2227–2237. ACL. Guangxu Xun, Kishlay Jha, Lu Su, and Jing Gao.
2018. Eann: Event adversarial neural networks for
Kashyap Popat. 2017. Assessing the credibility of multi-modal fake news detection. In Proceedings
claims on the web. In Proceedings of the 26th Inter- of the 24th acm sigkdd international conference on
national Conference on World Wide Web Compan- knowledge discovery & data mining, pages 849–857.
ion, pages 735–739. ACM.
Martin Potthast, Johannes Kiesel, Kevin Reinartz, Thomas Wolf, Lysandre Debut, Victor Sanh, Julien
Janek Bevendorff, and Benno Stein. 2017. A sty- Chaumond, Clement Delangue, Anthony Moi, Pier-
lometric inquiry into hyperpartisan and fake news. ric Cistac, Tim Rault, R’emi Louf, Morgan Funtow-
arXiv preprint arXiv:1702.05638. icz, and Jamie Brew. 2019. Huggingface’s trans-
formers: State-of-the-art natural language process-
Erzsébet Ravasz and Albert-László Barabási. 2003. Hi- ing. ArXiv, abs/1910.03771.
erarchical organization in complex networks. Physi-
cal review E, 67(2):026112. Ke Wu, Song Yang, and Kenny Q Zhu. 2015. False ru-
mors detection on sina weibo by propagation struc-
Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, tures. In 2015 IEEE 31st international conference
Rianne Van Den Berg, Ivan Titov, and Max Welling. on data engineering, pages 651–662. IEEE.
2018. Modeling relational data with graph convolu-
tional networks. In European Semantic Web Confer- John Zarocostas. 2020. How to fight an Infodemic.
ence, pages 593–607. Springer. The Lancet, 395(10255):676.
Xinyi Zhou, Jindi Wu, and Reza Zafarani. 2020. Safe:
Similarity-aware multi-modal fake news detection.
arXiv preprint arXiv:2003.04981.
Xinyi Zhou and Reza Zafarani. 2018. Fake news: A
survey of research, detection methods, and opportu-
nities. arXiv preprint arXiv:1812.00315.
A Appendix
A.1 Text preprocessing T rue P ositive
P recision =
T rue P ositive + F alse P ositive
We clean the raw text of the crawled articles of
the GossipCop dataset before using them for train-
ing. More specifically, we replace any URLs T rue P ositive
and hashtags in the text with the tokens [url] Recall =
T rue P ositive + F alse N egative
and [hashtag] respectively. We also replace
new line characters with a blank space and make A.5 Training Details
sure that class distributions across the train-val-test 1. To leverage effective batching of graph data
splits are the same. during training, we cluster the Graph into 300
dense sub-graphs using the METIS (Karypis
A.2 Hyper-parameters
and Kumar, 1998) graph clustering algorithm.
All our code is in PyTorch and we use the Hug- We then train all the GNN networks with a
gingFace library (Wolf et al., 2019) to train the batch-size of 16, ie, 16 of these sub-graphs are
transformer models. We grid-search over the fol- sampled at each pass as detailed in (Chiang
lowing values of the parameters for the respective et al., 2019). This vastly reduces the time,
models and choose the best setting based on best memory and computation complexity of large
F1 score on test set: sparse graphs.
1. CNN: learning rate = [5e-3, 1e-3, 5e-4, 1e-4], 2. Additionally, for GCN we adopt ”diagonal
dropout = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6], weight enhancement” by adding identity to the origi-
decay = [1e-3,2e-3] nal adjacency matrix A (Chiang et al., 2019)
and perform the normalization as:Ã = (D +
2. Transformers: learning rate = [5e-3, 1e-3, I)−1 (A + I).
5e-4, 1e-4], weight decay = [1e-3, 1e-2, 1e-
1, 5e-1], hidden dropout = [0.1, 0.2, 0.3, 0.4, 3. For SAGE we use ”mean” aggregation and
x0
0.5], attention dropout = [0.1, 0.2, 0.3, 0.4, normalize the output features as x0i where,
k i k2
0.5]
x0i is x0i = W1 xi + W2 · meanj∈N (i) xj .
3. GNNs: learning rate = [5e-3, 1e-3, 5e-4,
4. For GAT, we use 3 attention heads with atten-
1e-4], weight decay = [1e-3, 2e-3], hidden
tion dropout of 0.1 to stabilize training. We
dropout = [0.1, 0.2, 0.3, 0.4, 0.5], node mask
concatenate their linear combinations instead
= [0.1, 0.2, 0.3, 0.4, 0.5], hidden dimension =
of aggregating, to have a output of each layer
[128, 256, 512]
to be 3 × hidden dim.
The set of best hyper-parameters for all models A.6 Results with CNN text encoder
are reported in Table 3.
The results of the proposed SAFER framework
A.3 Hardware and Run Times with CNN used as the text-encoder are reported
in Table 5. We can note similar trends in the per-
We use NVIDIA Titanrtx 2080Ti for training
formance although the scores are slightly lower as
multiple-GPU models and 1080ti for single GPU
compared to GossipCop.
ones. In Table 4 we report the run times (per epoch)
for each model. A.7 Effect of graph sparsity and frequent
users
A.4 Evaluation Metric
In Table 6 we report the performance of all the
We use F1 score (of the target class, ie, fake class)
GNN variants of the proposed SAFER framework
to report all our performance. F1 is defined as :
for different subsets of highly active users.
P recision × Recall A.8 Community Graph
F1 = 2 ×
P recision + Recall
A portion of the community graph is visualized in
where, Precision and Recall are defined as: Figure 5.
 Graph Text
GCN GAT SAGE R - GCN R - GAT H y- GCN H y- GAT CNN R o BERT a
−4 −4 −4 −3 −4 −3 −3 −4
Learning rate 5 · 10 5 · 10 1 · 10 1 · 10 1 · 10 5 · 10 5 · 10 5 · 10 5 · 10−4
Weight Decay 1 · 10−3 1 · 10−3 2 · 10−3 2 · 10−3 1 · 10−3 1 · 10−3 1 · 10−3 1 · 10−3 5 · 10−1
Attention dropout NA 0.1 NA NA 0.1 NA NA NA 0.4
Hidden dropout 0.1 0.4 0.2 0.4 0.2 0.1 0.1 0.5 0.1
Node masking prob. 0.1 0.1 0.1 0.1 0.1 0.1 0.1 NA NA
Hidden dimension 256 512 128 512 512 256 512 384 1024
Learning rate 5 · 10−3 5 · 10−4 1 · 10−4 5 · 10−4 1 · 10−3 5 · 10−3 5 · 10−3 5 · 10−4 5 · 10−4
Weight Decay 1 · 10−3 2 · 10−3 2 · 10−3 2 · 10−3 1 · 10−3 1 · 10−3 1 · 10−3 1 · 10−3 5 · 10−1
Attention dropout NA NA NA NA NA NA NA NA 0.4
Hidden dropout 0.4 0.2 0.2 0.2 0.2 0.1 0.1 0.5 0.1
Node masking prob. 0.1 0.2 0.1 0.2 0.1 0.3 0.3 NA NA
Hidden dimension 512 512 128 512 512 512 256 384 1024

Table 3: Best Hyper-parameters for all the models on GossipCop (top) and HealthStory (bottom).

GossipCop HealthStory
Method No. of GPUs Run time (per epoch) No. of GPUs Run time (per epoch)
CNN 4 15 mins 4 1.25 mins
R o BERT a 4 6 mins 4 3 mins
SAGE 1 8.77 secs 1 1.49 secs
GCN 1 6.06 secs 1 1.91 secs
GAT 1 6.76 secs 1 1.96 secs
RGCN 1 6.92 secs 1 1.40 secs
RGAT 1 7.88 secs 1 2.16 secs
H y- GCN 1 10.39 secs 1 1.64 secs
H y- GAT 1 16.50 secs 1 2.97 secs

Table 4: Per epoch run times of all the models

4089 3504 3634

3142 3807 15143 3290
6094 3799 3959 3691 3996 2909
13647 3913 3936 3877
3637 3696 3860 3991 3004
142 3557 3112
1936 30 3930 3495 3063
1628
361 16749
43411 2823
15102 3724 15114 3347 3525
993 975 2908 4110 3360 2893
6670 1481 1126 1827 2109 1152 14622 3354 3397 4138 3823 15112 15192 3230
685 2993 3982 4315
1456 12111 6863 2876 15249 254 4211
352 2631 852 3847 3325 3340 3204 3981
3568
1114
2746
574
14710 1180 1938 6054 3221 7069 3704 15255 3163 15247
12455 2229 3318 3717 3892 4088
8509 15038 14602 2153 4162 4074 4153 4304
14822 14991 1542 14600 1364 3540 15160 3088 3066
14943 8653 4294 2199 4151
2767 739 14714 1375 3605 3570 3711 4075 15724 15141
1727 2431 2363 23016 4096 371
4121 1949 3205
2030 112 8934 2241 1366 3387 3720 3856 4085 3914 4181 3269
675 2530 3750 15109 4324
767 2760 14733 761 14550 14474 4115 4176
1159 4055 4171 3714 10752
15035 3501 3675
2556 1052 9276 9904 7166 25462 3683 27908 3032 5397
3606 3202
848 1029 2949 7300 13574
2678 14842 32358 2762 2923 15718 15158 15228 15262 3999
654 22293 790 941 2996 843 3867
2565 596 13 14827 16808 21443 14959 41153 3897
14840 16753 8044 3644
470 15021 11732 34814 44981 30355 46681 31154 392714528 3162 15730 2932
10278 12633
4086
286 23584 32439 7333 4072 3741
552 781 14749 9034 2093 34428 33375 443 2050 15207 3659
2665 581 41491 3430 201 15257 12165 3062 2905
1207 1550 2069 14788 29152 40429 2796 801 693 3270 3211 3716 12471
2464 30845 797
75 2296 1631 7296 346 1357 3187 2877 7194
2616 2233 2088 16858 7304 3480 3236 4109 4070 4305
133 9206 1358 43052 9271 3550 5020
175 26357 4378 15135
2426
14932 597 1424 2422 17040 1511 89 7279 4065
24976 25086 7289 4345 3233 4299 3078 4194
1225 2393 2586 8786 433 2696 1867 1584 7278 15775 7273 4243 3885 3214
281 7361 3268 3929 3259
36097 14890 2019 3783 6011 4349
465 14562 7205 3849 15094
2669 2104 1747 11499 16760 10 3643 4023 2985 4362 4184 3803
14608 836 7852 34985 5722 2907
14673 1107 3977 3363 3190
2524 1646 12527 14949 3144 3218 3283 15120 15239 3676
845 14866 3024
14864 888 2261 1624 3932 3009 3919 4148 4025
15231
7332
9469
1890 39789 3093
19043
1303
23735
29119

4382 14995
13345 39087

5737 14947 2205

10855 5789
5563 6981
849 15189 11422 9305
1371 6380 7163

Figure 5: Visualization of a small portion of the fake news community graph. Green nodes represent the articles of the dataset
while red nodes represent users that shared them.

A.9 t-SNE visualizations this article, we see that on average these users
shared 5.8 fake articles while just 0.45 real ones
A.10 Qualitative Analysis (13 times more likely to share fake content than
We assess the performance of the SAFER (GCN) real), strongly indicating that the community of
variant on Gossipcop in Figure 7a. We see that users that are involved in sharing of this article are
the first article is a fake article which RoBERTa responsible for propagation of fake news. Taking
incorrectly classifies as real. However, looking at this strong community-based information into con-
the content-sharing behavior of users that shared sideration, SAFER is able to correctly classify this
 Model GossipCop HealthStory
†
HAN 67.20 - 20

dEFEND† 75.00 - 10

Text SAFE
‡
89.50 - 0

20 10
15
10

CNN 66.73 53.81 5

0
5 20
20

R o BERT a 68.55 57.54 10

15
20
10
0
10

Maj. sharing baseline 77.19 8.20 (a)

SAGE 91.11 56.34
GCN 91.95 56.84
GAT 92.41 56.91
SAFER R - GCN 93.48 60.45 10
15

R - GAT 93.75 61.58 0

5
5

H y- GCN 92.34 59.75 10

15

H y- GAT 91.56 59.89 20

20
20

10 10

0 0
10
10
20

Table 5: F1 scores (fake class) on GossipCop and HealthStory 20

† (b)
using CNN as the text encoder. ( ) denotes results reported
from Shu et al. (2019a) and (‡ ) from Zhou et al. (2020). Bold
20

figure denotes significantly better than other methods for that 10

dataset. Underscore figures denote significantly better scores 0

than baselines but not significantly different from each other. 10

20

20
15
10
5
0

Setting ρ SAFER F1 20 10 0 10 20 15
10
5

SAGE 82.14 (c)

GCN 81.40
GAT 81.01 Figure 6: 3-D t-SNE plots for representations of test articles
excl.>10% 0.78 R - GCN 81.87 produced by (a) SAFER(GAT) (b) SAFER(GCN), and (c)
R - GAT 82.27 SAFER(RGCN). Red dots denote fake articles.
H y- GCN 82.16
H y- GAT 81.81
SAGE 76.96 tures indicate that the users interacting with the
GCN 76.87
GAT 77.07 article share 16.2 fake articles and 7.8 real ones
excl.>5% 0.71 R - GCN 77.16
R - GAT 77.32 on average (2.1 times more likely to share fake).
H y- GCN 77.13 SAFER takes this information into account and
H y- GAT 77.01
classifies it correctly as fake. Similarly for the sec-
SAGE 69.52 ond article, the interacting users share 40 real and
GCN 69.14
GAT 68.67 19.96 fake articles on average (2 times more likely
excl.>1% 0.65 R - GCN 65.89 to share real) which helps the proposed method to
R - GAT 65.32
H y- GCN 71.99 correctly classify it as real.
H y- GAT 72.05

Table 6: Results of SAFER variants on varying subsets of

user nodes on GossipCop. ρ denotes relative graph density.

article as fake. Similarly, the second article is a real

article which is misclassified as fake by RoBERTa
by looking at the text alone. However, the GNN
features show that the users that shared this article
have on average shared 533 real articles and 96.7
fake ones (5.5 times more likely to share a real
article than a fake one). This is taken as a strong
signal that the users are reliable and do not engage
in malicious sharing of content. SAFER is then
able to correctly classify this article as real.
We observe similar behavior of the models on
HealthStory in Figure 7b. The first article is mis-
classified as real by RoBERTa but the GNN fea-
 Text GNN RoBERTa SAFER(GCN) Ground Truth

The latest Bacon question: Is he really singing the praises of

President Donald Trump? In the depressing age of fake news,
there is a dual responsibility to identify the difference between
news that is actually false or fabricated and news that simply rubs ?
us the wrong way.... a recent story claiming Kevin Bacon believes
we are all "blessed" to have Donald Trump as our president is
utterly untrue. The dubious news blog Worldwide Politics
published a story on Sunday.....Worldwide Politics is the epitome
of a fake news website.... ran a made-up story claiming actress
Julia Roberts called Trump "the best president in US history......

Surprise! It was widely believed that Jason Momoa and Lisa

Lilakoi Bonet married in November 2007, but they didnt make it
official until this October...Jason Momoa and Lisa Bonet had an
official wedding a few weeks ago at their house in Topanga, ?
California one insider reveals.....A second source confirms to Us
....according to a Facebook user the Game of Thrones alum wore
an unbuttoned white shirt and a lei. He styled his long locks into a
man bun. The Cosby Show alum looked beautiful.... some of the
guests performed a traditional Mori haka for the bride....

(a)

Text GNN RoBERTa SAFER (GCN) Ground Truth

The plant extract resveratrol, found in the skin of red grapes, appears
to suppress inflammation and may fight aging in humans, according to
a new study.... apparently because resveratrol affects a gene
associated with longevity....they have found that resveratrol reduces
?
inflammation in humans that could lead to heart disease, stroke, and
type 2 diabetes.... 20 people and put them at random into two groups,
one receiving a placebo and the other a supplement containing 40
milligrams of resveratrol...fasting blood samples were taken at the start
of the trial and then at intervals of one, three, and six weeks...people
taking resveratrol also showed suppression of the.. TNF...blood
samples from those on placebo showed no significant change in pro-
inflammatory markers.

A widely prescribed drug used to shrink enlarged prostates

appears to reduce the incidence of prostate cancer in men with an
increased risk for the disease. In a trial involving more than 8,000
men from 42 countries, those who took the drug Avodart had a
23% lower risk of being diagnosed with prostate cancer over four ?
years of treatment, compared to men who did not take the drug....a
total of 8,231 men between the ages of 50 and 75 took part in the
study, published in the New England Journal of Medicine.....the
men were assigned to receive either placebo treatments or daily
0.5-milligram doses of Avodart for four years.....

(b)

Figure 7: Demonstrating the importance of community-based features of the proposed method on (a) Gossipcop and, (b)
HealthStory. Text in red denotes a fake article, while in green denotes a real one. Black central node denotes the target article
node that we are trying to classify, blue nodes denote the users that shared this article while red and green nodes denote the other
fake and real articles these users have interacted with respectively. Predictions by different models stated on the right.