You are on page 1of 5

2022 IEEE 2nd International Conference on Power, Electronics and Computer Applications (ICPECA)

A Multilevel Graph Convolution Neural Network


Model for Rumor Detection
2022 IEEE 2nd International Conference on Power, Electronics and Computer Applications (ICPECA) | 978-1-6654-4276-3/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICPECA53709.2022.9719043

Yuanyuan Ma Shouzhi Xu Fangmin Dong


College of Computer and Information College of Computer and Information College of Computer and Information
Technology, Technology, Technology,
China Three Gorges University, China Three Gorges University, China Three Gorges University,
Yichang, China Yichang, China Yichang, China
mayuanyuan@ctgu.edu.cn xsz@ctgu.edu.cn fmdong@ctgu.edu.cn

Abstract—Rumor detection is a challenging task on social multilevel graph convolution network model (MGCN)
medias. When a post is propagated on social media, it usually including post level and event level for rumor detection is
contains four types of information: 1) content; 2) time of proposed in this paper. In this model, the post level social
publishing; 3) structure of propagation; 4) social interaction. network graph is the propagation graph, and it takes the time
In most previous studies, the information has not been difference after publishing as the edge weight. Event level
effectively combined to detect rumors. A multilevel graph social network graph takes events as nodes and the
convolution model including post level and event level is correlation degree between events as edge weight. The GCN
proposed to detect rumors in this paper. For post level graph network is applied to aggregate the post text features and
convolution network based on propagation relationship, it uses
propagation clues. Next the aggregation representation is
a graph convolution network with rumor propagation graph to
learn post level features. For event level graph convolution
spliced with the source post text features to obtain the post
based on event interaction relationship, a graph convolution level fusion features. Then the GCN network is used to
network with event relationship graph is applied to bridge post bridge the post level fusion features and event interaction
level features and event interaction information to obtain the clues as the final fusion multi-level features of the event
feature representation of events. The experiment results shows node. The main points of this paper are as follows:
that rumor detection accuracy of our model is 94.3%, which is 1) GCN is introduced for rumor detection. The MGCN
superior to other newly models.
model considers both post level social networks and event
Keywords—Rumor Detection, Graph Convolution Neural
level social networks. The deep integration of two levels of
Network, Propagation Structure, Social Networks social networks is used to detect rumors.
2) The source post text feature and the post level feature
I. INTRODUCTION are combined to improve rumor detection.
Recently, with popularization of social media such as
Microblog and twitter, people can freely release information II. RELATED WORK
and directly establish contact and communication, which not A. The Conventional Methods
only accelerates the flow and dissemination of public
information, but also makes generate and disseminate rumor Most of early researches on rumor detection focused on
events more easily and fastly. And most of the information in artificial design features. Castillo[1] designed a variety of
social media is released by individuals which are lack of features from a statistical point of view, such as text length,
supervision. Some people even deliberately issue false number of emotional words, number of user registration days
information for attention or other special purposes. What's and number of friends, and used the decision-tree
more, Rumor events are often inflammatory, since that rumor classification algorithm to evaluate the authenticity of
publishers may consciously use some inflammatory information related to specific topics on the twitter platform.
statements in order to make their rumors spread in a wider Wu[2] considered that the traditional rumor detection
range. Users who comment on and forward rumor are usually method ignored the post propagation structure. They
more vulnerable to the temptation of stimulating statements, proposed a hybrid SVM based on a graph kernel to capture
and may participate in the spread of more rumor events. the semantic features and high-order propagation patterns to
improve detection accuracy. Ma[3] proposed a kernel-based
Most of previous methods regard the rumor event as an approach to capture high-level representations that
independent individual, only consider the characteristics of distinguish different types of rumors by assessing the
the event itself (such as propagation structure, post sequence similarity between propagation trees. These conventional
feature and user feature), and ignore the interaction between methods rely heavily on handcrafted feature engineering to
events on the social platform. In fact, events are not extract informative feature , which is inefficient and usually
independent on social platforms. Every event in the social require high processing cost.
platform and the users associated with the event are related,
including posting, forwarding comments, etc. And the B. Deep Learning Methods
interaction between these users is also a reflection of the The neural network can efficiently explore the features of
relationship between events related to these users. events on social medias, so the direction of rumor detection
has gradually shifted from feature mining to model structure
In order to consider propagation feature of event and the
design. Ma[4] discovered the continuity of the text stream,
interaction information on social platforms adequately, a

978-1-6654-4276-3/22/$31.00 ©2022 IEEE 1225 21–23 January 2022, Shenyang, China

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on October 21,2022 at 02:43:55 UTC from IEEE Xplore. Restrictions apply.
and that RNN could capture the dynamic time signals of j  [1, mi ] , its text feature vector is x ij . In addition, each
rumor forwarding. So a RNN based rumor detection model
uses timing relationship of relevant posts to learn the event ci is related to a label yi  {F , T } (i.e., False
semantic features of tweet context over time. In addition, She Rumor or True Rumor).
built top-down and bottom-up tree-structured RvNN to get gi  (Vi , Ei ) : Node set
1) For post level graph
the hidden feature vectors from both text contents and
propagation sequences[5]. Liu[6] proposed a CNN+RNN Vi  {c , c , c ,..., c } represents all posts (source post and
i
0
i
1
i
2
i
mi
model based on time series classifier to capture the global response posts) nodes of the event ci , and edge set
and local changes of user characteristics along the
propagation path to detect false news. Its input is time-series Ei  {esti | s, t  0,1,..., mi } represents the set of edges from
in news forwarding paths. Alkhodairet[7] trained a LSTM response post to the forwarded post. For example, if cti is a
model which only learns the source post’s content to classify
rumors. Chen[8] provide an RNN-based deep attention response to csi , there will be an edge between cti and csi ,
model, which determined different features by learning latent i.e., esti . Denote Ai  R( mi 1)( mi 1) as the adjacency matrix
representation from sequential tweets. However, the above
of g i where
methods only care the order of post publishing, so they can’t
learn the structural characteristics of rumor propagation. It
needs introduce GCN to model the propagation relationship.  1 i
1  (ln time ) /10 , if est exists
C. The GCN Method  as ,t  Ai , as ,t   s ,t 
Graph neural network is an effective framework for 0 , otherwise

learning graph structure representation, which is mainly used
for graph node classification and graph structure
among them, esti  Ei , times ,t represents the time
classification. Yu[9] proposed a GCN model integrating
static features such as user basic information and text content difference between the publishing time of forwarded post
to capture the propagation characteristics of rumors and csi and response post cti , as ,t represents the weight of
detect rumors. Xu[10] proposed a Dual-grained Feature edges between nodes and it also represents the strength of
Aggregation graph neural networks model. It applied a GCN relationship between nodes, which is equivalent to the
model with a graph of rumor propagation to learn the text- impact of response posts on the forwarded post.
granularity representations with the spreading of events and
employed a GNN with a document graph to update 2) For event level graph G  ( Node, Edge) : It is
aggregated features of both word and text granularity. Wu[11] structured according to the co-participating users among the
proposed a gated graph neural network based algorithm events, where node set Node  {c1 , c2 ,..., cn } , and edge set
called PGNN, which can generate powerful representations Edge  {edge pq | p, q  1, 2,..., n} represents the connection
for each node in the propagation graph. This model
dynamically adjusted the weight of each node in the between events. For example, if there are any co-
propagation graph by introducing an attention mechanism. participating users between event c p and event cq , there
Graph neural network methods make full use of the event will be an edge between them, i.e., edge pq . Denote
propagation structure, but these methods treat event as an
isolated individual, ignoring the interaction information A  R n n as the adjacency matrix of G , where
between events on social networks.

un p , q , if edgepq exists
In our opinion, the propagation structure of event and the  ap , q  A, ap , q   
interactive information between events are both important 
0 , otherwise
parts. Inspired by the GCN, MGCN model is proposed to
capture and bridge both post-level and event-level features of
among them, edgepq  Edge , un p , q represents the
rumor detection events.
correlation degree between events, which is the number of
III. STATEMENT OF RUMOR STRUCTURES users who participate in both event c p and event c p at
A. Basic Symbols the same time, a p , q represents the weight of the edges
For a rumor detection dataset of events, each event ci between nodes, which represents the strength of the
relationship between event nodes.
consists of a series of related posts, and each event
corresponds to a propagation graph g i . All events form an 3) For the GCN Method: GCN[12] is a multi-layer
event diagram G, which can be given by, neural network, which processes graph data and generates
embedded vectors of nodes in a rumor event according to
 C  {c1 , c2 ,..., cn , G}  their neighborhoods. GCN can encode local graph structure
and node features. Its forward propagation formula is given
by,
 ci  {c0i , c1i , c2i ,..., cmi i , gi } 
 ˆ (l )W (l ) ) 
H (l 1)  F ( A, H (l ) ;W (l ) )  ( AH
where n is the number of events, mi is the number of
response posts for the event ci , c0i is the source post, its where   D1/ 2 AD1/ 2 is the normalized symmetric
i i
text feature vector is x . c refers to the j-th response post adjacency matrix, D is the degree matrix and W (l )
0 j

1226

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on October 21,2022 at 02:43:55 UTC from IEEE Xplore. Restrictions apply.
represents the trainable weight matrix for the (l+1)-th obtain the final representation of events for rumor detection.
hidden layer. ( ) denotes an activation function. The MGCN model consists of three modules: post level
GCN module, event level GCN module and rumor detection
Then the three-layer GCN model function is as follows, module, as shown in Fig. 1.
1) The post level GCN module constructs the post level
 H (3)  ( Aˆ ( Aˆ ( AH
ˆ (0)W (0) )W (1) )W (2) ) 
social network graph, and applies GCN network to aggregate
the post text features and propagation clues. Then it splices
Here, W (0) is an input-to-hidden weight matrix for the the aggregation representation with the source post text
first hidden layer. W (1) is a hidden weight matrix, while features to obtain the post level fusion features.
W (2) is a hidden-to-output weight matrix. 2) The event level GCN module builds the event
interaction graph, and uses the GCN network to bridge the
B. Model Framework post level fusion features and event interaction clues as the
A multilevel graph convolution network model for rumor final fusion multi-level features of the event node.
detection (MGCN, in brief) is proposed. The core idea of this
model is to use graph convolution network to obtain and fuse 3) The rumor detection module trains the classification
the information of two different levels of social graph to function to predict the label of events.

Fig. 1. MGCN Model

C. Post Level GCN Module of PLGCN. We employ mean-pooling operators and max-
pooling operators to aggregate information from these node
The propagation of information between the source post representations as Eq. (8). The propagation features
and the reply post cannot be looked down upon. We use the
{hgi | i  1, 2,..., n} of all events constitute the propagation
propagation structure and text feature to generate event
individual feature to improve the performance of rumor feature matrix H g  [hg1 , hg2 , hg3 ,..., hgn ] .
T

detection. After constructing the post level social graph, as


well as the event propagation graph gi  (Vi , Ei ) , by Then, the source-post text feature matrix
substituting Ai  R( mi 1)( mi 1) and X i  [ x0 , x1 , x2 ,..., xmi ]
i i i i T X   [ x10 , x02 , x03 ,..., x0n ]T is concatenated with the
propagation feature matrix H g  [hg1 , hg2 , hg3 ,..., hgn ]
T
to Eq. (6) over three layers, we write the equations for Post to
Level GCN (PLGCN) as below: construct a fused feature matrix F  [ f1 , f 2 , f3 ,..., f n ]T
, as
Eq. (9):
 Hi(3)  ( Aˆi ( Aˆi ( Aˆi X iW (0) )W (1) )W (2) ) 
 F  cat( X , H g ) 
 hgi  MeanPooling( H i(3) )  MaxPooling( H i(3) ) 
D. Event Level GCN Module
( mi 1)v3
where H i
(3)
R represents the output-layer If only considering the social network at the post level
and ignoring the social network at the event level, it may
feature matrix in the three-layer PLGCN. Wi (0)  R d v1 , affect the detection results. Therefore, the Event Level GCN
Wi (1)  Rv1 v2 , Wi (2)  Rv2 v3 are the trainable weight matrix module (ELGCN, in brief) is proposed, which uses the

1227

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on October 21,2022 at 02:43:55 UTC from IEEE Xplore. Restrictions apply.
relationship between events to establish the event level 1) DTR [16]: a decision-tree-based ranking model to
social graph. It aggregates the fusion feature of post level identify trending rumors through ranking the clustered
and the interaction clues of event level social graph, and disputed factual claims based on statistical features.
improves the performance of rumor detection. After
2) SVM-RBF [17]: a rumor detecting method using
constructing the event level social graph G  ( Node, Edge) , manual features and SVM classifier with RBF kernel
by substituting A and F to Eq. (6) over three layers, the function. The hand-worked features are extracted from
equations for ELGCN are given as fellows: Weibo.
3) RvNN [5]: a rumor detecting approach based on tree-
 H (3)  ( Aˆ ( Aˆ ( AFW
ˆ (0)
)W (1) )W (2) )  structured recursive neural networks. It learns post
representations via event propagation.
where H (3)  Rnv6 represent the output-layer feature
4) PPC_RNN+CNN [6]: an early detection approach of
matrix of three-layer ELGCN. W (0)  R ( d  v3 )v4 , fake news through classifying propagation paths. They
W (1)  R ( d  v4 )v5 and W (2)  R ( d  v5 )v6 are the trainable construct the paths as multivariate time series and build a
weight matrixes of ELGCN. time series classifier incorporating RNN and CNN.
E. Rumor Detection Module B. Result Analysis
After the above model, we get the updated node Experimental Result: from Table II, it shows that the
characteristic matrix H (3)  [h1 , h2 , h3 ,..., hn ]T . The label yˆi rumor detection model proposed by our research is better
than traditional classifiers and the other classifiers based on
the event is calculated via several full connection layers and other neural networks. In terms of accuracy, the accuracy of
a softmax layer: our model is 12.5%, 3.5% and 2.7% higher than SVM-RBF,
RvNN and PPC_RNN+CNN respectively on the Weibo
 yˆi  softmax (FC (hi ))  dataset. The F1 values of positive samples and negative
samples were 2.3% and 3% higher than those of RvNN
 L   yi log( yˆi )  respectively. The Recall of positive samples and the Recall
i of negative samples are 2.3% and 3% higher than those of
RvNN respectively. The increase of Recall shows that this
where yˆi  R1C is the probability vector for all the method is more sensitive to rumor events, it can reduce the
misjudgment rate of rumor events, find more rumor events to
classes used to predict the label of event ci , C is the reduce the adverse effects of rumor events. Obviously, the
number of categories, hi is the updated event node multilevel graph convolution neural network model
representation, yi is the real label and L is the cross including post level social graph and event level social graph
can improve the classification ability of the model to rumor
entropy loss function. information to a certain extent.
IV. EXPERIMENTS AND ANALYSIS
TABLE II. RUMOR DETECTION RESULTS
A. Dataset and Experiment Settings Method Class Acc. Prec. Recall F1
We evaluate the proposed method on the real-world DTR
F
0.732
0.738 0.715 0.726
datasets: Weibo[13]. Weibo is the most popular social media T 0.726 0.749 0.737
sites in China. It includes two categories of labels, which are F 0.822 0.812 0.817
SVM-RBF 0.818
False Rumor (F) and True Rumor (T). The statistics of the T 0.815 0.824 0.819
F 0.912 0.897 0.905
dataset are shown in Table I. The Doc2vec model [14] is RvNN
T
0.908
0.904 0.918 0.911
used to generate the 300-dimensional text vector of the post. PPC_RNN F 0.884 0.957 0.919
In the PLGCN module, the hidden feature vector dimension 0.916
+CNN T 0.955 0.876 0.913
is 200-dimensional, and the output vector dimension is 100- F 0.956 0.929 0.942
MGCN 0.943
dimensional. In the ELGCN module, the hidden feature T 0.930 0.956 0.943
vector dimension is 500 dimensions, and the output vector
dimension is 64 dimensions. The rate of dropout is 0.3 and Ablation Study: To analyze the effect of each variant in
the learning rate is 0.001. Moreover, the Adam algorithm is MGCN on rumor detection, a set of ablation experiments is
applied to optimize the model. The loss function adopts cross conducted. The experimental results are shown in Fig 2.
entropy loss function. The training process is iterated upon UPUE represents a model that uses unweighted post level
200 epochs, and early stopping[15] is applied when the graph and unweighted event level graph, WPUE represents
validation loss stops decreasing by 10 epochs. the model which uses weighted post level graph and
unweighted event level graph, UPWE represents the model
TABLE I. STATISTICS OF THE DATASET which uses unweighted post level graph and weighted event
level graph, PLGCN indicates that only the Post Level graph
Statistic Weibo convolution model is used, and ELGCN indicates that only
# of Events 4664 the Event Level graph convolution model is used. The
# of Rumors 2351
# of Non-rumors 2313
accuracy is only 87% when using two unweighted graphs.
# of Posts 3,805,656 Adding weighted post level graph and weighted event level
# of Users 2,746,818 graph respectively, the accuracy is improved by 4% and
2.2% respectively. It can be concluded that the effect of
We compare the proposed method with some state-of- rumor detection can be greatly improved by the different
the-art methods, including: importance of response posts and the relevance of events.

1228

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on October 21,2022 at 02:43:55 UTC from IEEE Xplore. Restrictions apply.
The accuracy of PLGCN model is only 76%, and the 651–662.
performance of the original model is severely degraded. The [3] J. Ma, W. Gao, K. Wong, “Detect rumors in microblog posts using
accuracy of ELGCN model can reach 91%, but it is still propagation structure via kernel learning,” in the 55th Annual
Meeting of the Association for Computational Linguistics, ACL, 2017,
2.7% lower than MGCN. It can be inferred that both PLGCN pp. 708–717.
and ELGCN are inseparable in our model, and each plays an [4] J. Ma et al. “Detecting rumors from microblogs with recurrent neural
important role in MGCN model. It shows that the fusion networks,” in the 25th International Joint Conference on Artificial
features of multilevel graph convolution neural network help Intelligence, Palo Alto: AAAI Press, 2016, pp. 3818–3824.
to improve the performance of the model. [5] J. Ma, W. Gao, K. Wong, “Rumor detection on twitter with tree-
structured recursive neural networks,” in the 56th Annual Meeting of
the Association for Computational Linguistics, Stroudsburg: ACL,
2018, pp. 1980–1989.
[6] Y. Liu, Y.B. Wu, “Early detection of fake news on social media
through propagation path classification with recurrent and
convolutional networks,” in the 32th AAAI Conference on Artificial
Intelligence, pp. 354–361. AAAI Press (2018).
[7] S. A. Alkhodair, S. H. Ding, B. C. Fung, J. Liu, “Detecting breaking
news rumors of emerging topics in social media,” Information
Processing & Management, vol. 57(2), 102018, March 2020.
[8] T. Chen, X. Li, H. Yin, J. Zhang, “Call attention to rumors: Deep
attention based recurrent neural networks for early rumor detection,”
Fig. 2. The results of ablation experiment. In Pacific-Asia conference on knowledge discovery and data mining,
Cham: Springer , 2018, pp. 40-52.
V. CONCLUSION [9] K. Yu, H. Jiang, T. Li, S. Han, X. Wu, “Data fusion oriented graph
convolution network model for rumor detection,” IEEE Transactions
The proposed MGCN model for rumor detection uses a on Network and Service Management, vol. 17(4), pp. 2171–2181,
multilevel graph convolution model including post level and December 2020.
event level is proposed to detect rumors. For post level graph [10] S. Xu et al. “Rumor Detection on Microblogs Using Dual-Grained
Feature via Graph Neural Networks,” In Pacific Rim International
convolution network based on propagation relationship, it Conference on Artificial Intelligence, Cham: Springer, 2021, pp. 205-
uses a graph convolution network with rumor propagation 216.
graph to learn post level feature. For event level graph [11] Z. Wu, D. Pi, J. Chen, M. Xie, J. Cao, “Rumor detection based on
convolution based on event interaction relationship, a graph propagation graph neural network with attention mechanism,” Expert
convolution network with event relationship graph is applied systems with applications, vol. 158, 113595, November 2020.
to bridge post level feature and event interaction information [12] M. Defferrard, X. Bresson, P. Vandergheynst, “Convolutional neural
to obtain the feature representation of events. The results of networks on graphs with fast localized spectral filtering,” in the 30th
the experiments on the Sina Weibo show that this method International Conference on Neural Information Processing Systems,
Red Hook: Curran Associates Inc, 2016, pp. 3837-3845.
can improve the accuracy of the model compared with the
[13] J. Ma, W. Gao, Z. Wei, Y. Lu, K.F. Wong, “Detect rumors using time
baseline method. In the future, our next research step will series of social context information on microblogging websites,” in
focus on improving the feature fusion and GCN the 24th ACM International on Conference on Information and
representation models to achieve more efficient and accurate Knowledge Management, New York: ACM, 2015, pp. 1751–1754.
rumor detection. [14] Q. Le, Tomas. Mikolov, “Distributed Representations of Sentences
and Documents,” in the 31st International Conference on International
ACKNOWLEDGMENT Conference on Machine Learning, 2014, pp. 1188-1196.
[15] Y. Yao, L. Rosasco, A. Caponnetto, “On early stopping in gradient
This work was supported in part by the National Natural descent learning,” Constructive Approximation, vol. 26(2), pp. 289–
Science Foundation of China (Grant No. U1703261 ). The 315, April 2007.
corresponding author is Shouzhi Xu. [16] Z. Zhao, P. Resnick, Q. Mei, “Enquiring minds: Early detection of
rumors in social media from enquiry posts,” in the 24th International
REFERENCES Conference on World Wide Web, Republic and Canton of Geneva,
[1] C. Castillo, M. Mendoza, B. Poblete, “Information credibility on CHE: International World Wide Web Conferences Steering
twitter,” in the 20th international conference on World Wide Web, Committee, 2015, pp. 1395–1405.
New York: ACM, 2011, pp. 675–684. [17] F. Yang, Y.Liu, X. Yu, M. Yang, “Automatic detection of rumor on
[2] K. Wu, S. Yang, K.Q. Zhu, “False rumors detection on sina weibo by sina weibo,” in the 18th ACM SIGKDD International Conference on
propagation structures,” in the IEEE 31st International Conference on Knowledge Discovery and Data Mining, New York: ACM, 2012, pp.
Data Engineering, Los Alamitos IEEE: Computer Society, 2015, pp. 1–7.

1229

Authorized licensed use limited to: Universidad Nacional Autonoma De Mexico (UNAM). Downloaded on October 21,2022 at 02:43:55 UTC from IEEE Xplore. Restrictions apply.

You might also like