Detecting fake news as early as possible has attracted growing attention due to its fast-spreading nature and the significant harm it can cause. As demonstrated in recent studies, the propagation pattern of fake news on social media difers from that of real news, and a number of works have extracted diferent types of features from the propagation pattern for detection. However, a major limitation of this approach is that the propagation network is not fully available in the early stages, and may take a long time to complete. As a result, existing network-based fake news detection methods yield low accuracy during the early stages of propagation. To bridge the research gap, in this work we: (1) propose a novel network embedding algorithm, based on the investigation of a wide range of features obtained from the propagation network, which are not well studied in previous work; and (2) design an autoencoder-based neural architecture to predict the embedding of the complete propagation network using the partially available network in the early stages of propagation. Our experiments show that with the predicted embedding for the complete propagation network, our model can achieve state-of-the-art performance while only having access to the early stage propagation network.
detection [
It has been demonstrated that the propagation pattern While the growing popularity of social media has greatly of news on social media, e.g., tweets and retweets of news facilitated the exchange of information, it also provides on Twitter, can facilitate the detection of fake news [7, an ideal platform to spread fake news, especially inten- 8, 9, 10, 11], since the propagation pattern of fake news tional disinformation, which has already and will con- exhibits distinctive characteristics. However, instead of tinue to cause significant damage. relying on the entire propagation network, which may
Even though a number of independent fact-checking take days or even weeks to complete, we only use the
organisations have emerged globally over recent years, initial network that, for instance, belongs to the first 100
the sheer volume of fake news makes it infeasible to tweets, or tweets posted within the first few hours, to
rely entirely on human investigation. In addition, what verify a news item. Specifically, the main contributions
makes the task even more challenging is that fake news of this work include (Figure 1 provides an overview):
needs to be detected at an early stage before it becomes • We investigate a range of local and global features of
widespread, since it is dificult to correct people’s per- the propagation network, including temporal-based,
textception towards an issue once it is formed, even if the based and user-based, and compare their contributions
previous impression is inaccurate [
The remainder of this paper is organised as follows: Section 2 defines the problem of fake news early detection; Section 3 describes how to embed the propagation network; Section 4 introduces the network embeddingbased detection algorithm; Section 5 provides the experimental verification of the designed algorithm; Section 6
Local Network
Embedding
Detection Deadline Tweets Posted within the Detection Deadline Tweets Posted after the Detection Deadline Source Node/ News Record Node-level Aggregation
Autoencoder
Fake News Detector
Binary
Classifier
Embedding of the Global Network Initial Propagation
Embedding Network
Predicted Embedding
of the Complete Propagation Network reviews previous work on fake news detection; and finally Section 7 concludes the paper and ofers directions for future work.
We define the problem of fake news early detection as follows: let be a set of labelled news records. Each record ∈ is represented as a tuple ⟨ , , , , ⟩, where (1) is the timestamp when is published online; (2) is the text content of ; (3) is the propagation network of at timestamp + (further explained below); (4) is the set of timeline tweets posted by the users involved in , i.e., it provides background information of the news spreaders. Note that does not necessarily always have to contain the latest timeline tweets; and (5) is the label: is 1 if is false and 0 otherwise.
Each propagation network is an attributed directed graph ( , , ), where: • is the set of vertices/nodes, and each node is a tweet/retweet with the corresponding user. A special case is that an extra node representing the news is added to link the network together—it is called the source node hereafter. • is the set of edges. Here, edges represent how a news item spreads from one person to another as shown in Fig 1. However, Twitter APIs do not provide the immediate source of a retweet. To solve this problem, within each cascade we first sort the tweets by their timestamps, and then search for the potential source of a retweet from all the tweets that are published earlier. Specifically, there is an edge from node to node if (1) the user of tweet mentions the user of tweet ; or (2) tweet is public and tweet is posted within a certain period of time after tweet , e.g., five hours. The follower and following relations are not included when constructing the edges of the propagation network, as they may not be available in real time due to the much stricter rate limit of the corresponding Twitter APIs, which prohibits the timely detection of fake news. • is the set of node-level and network-level features for , which are explained in detail in Section 3.
The problem is to predict the label for unlabelled news records ∈ as false or real news records within a detection deadline Δ, where for ∈ is only available for ≤ Δ.
In this section, we propose a simple yet efective unsupervised network representation learning method to embed the propagation network. Formally, for a given propagation network of news record at timestamp + , representation learning aims to learn a mapping function : → ℎ ∈ R such that the obtained embedding ℎ is useful for predicting the label of the news record. Moreover, we analyse the informativeness of the learned embeddings for the initial propagation network at the detection deadline Δ 1 and for the complete propagation network
( >> Δ).
Datasets. We conduct all our experiments on two
publicly available datasets introduced in [
Δ as the initial propagation network.
2https://www.politifact.com/ 3https://www.gossipcop.com/ sophisticated neural architectures such as graph neural networks [14] that use data-driven trainable kernels to perform node-level aggregation, our approach is more straightforward and hence easier to interpret. 3.1.2. Node-level Features
Input: propagation network = ( , , )
source node of ∈
gamma ∈ [
Output: The local representation ( ) 1 ℎ0 ← ∀v∈ 2 for in 1, 2, ..., do 3 for in do ∑︁∀(, )∈ ℎ−1 ℎ ← ℎ−1 + (1 − ) ∑︁ ∀(, )∈ 1 4 5 6 end
end
7 ( ) ← ℎ
8 ( )
news records and all the tweets and retweets for each
news item. Please refer to [
3.1.1. Node-level Feature Aggregation (1) user-based features; (2) text-based features; and (3)
The nodes in a propagation network typically have com- temporal features.
plex multi-modal attributes e.g., temporal-based, text- Node-level User-based Features. The following
userbased and user-based, which are useful to characterize based features are studied in our experiments: whether
the propagation network. Previous work [
The proposed approach iteratively updates the embed- can be good indicators to identify less credible users.
ding of the nodes based on their one-hop neighbours. In addition, the finding in [
0 2 4
8 6 which may be due to the domain diference of the two Node-level Temporal Features. Moreover, we analyse datasets. the following node-level temporal features to further
Another interesting observation is that the correlation capture the diference in the dissemination between fake of each feature converges after a few iterations (≈ 8) us- and real news: the time diference with the source node ing the proposed node-level aggregation approach. This (15); the time diference with the immediate predecessor observation indicates that the nodes that are close to the (16); the average time diference with the immediate source node are more informative compared to the rest successors (17); user account timestamp (18). in the propagation networks. According to the correlation analysis in Figure 4, the selected features show moderate correlations with the Node-level Text-based Features. We further study news labels for both datasets, and the results are also text-based features as listed below: the sentiment scores more consistent over diferent values of , compared computed using VADER4 using text content in the tweets with the other node-level features. (7); the frequency of positive words (8); the frequency In summary, the proposed label propagation scheme of negative words (9); the number of emojis (10); the can capture up to -hop neighbour information to genernumber of mentions (11); the number of hashtags (12); ate the embedding for the source node based on the nodeand the percentage of tweets related to the target topic level features. Our empirical analysis shows that the (13)—we collect the timeline tweets for each user, and nodes in close vicinity to the source node are mostly inrun tweet topic classification. For the dataset of PolitiFact formative to generate useful local representations. Thus (or GossipCop), we calculate the percentage of tweets the proposed label propagation scheme with a limited whose topic is classified as “politics" (or “entertainment"). value is suficient for performing node-level feature
The node-level text features can be helpful to under- aggregation.
stand the linguistic diferences of the text contents
generated by the users engaging with fake news and real 3.2. Global Representation
news. As shown in Figure 3, for both datasets a subset of
the above features show relatively high correlation with In addition to local features, the following network-level
the news labels, e.g., features 9, 11 for PolitiFact, and features are also extracted to represent the structural
features 11, 13 for GossipCop. This aligns with well- properties of each network , which is denoted as the
defined theories—for example, it has been demonstrated global representation ( ) of the network.
that user-bias is a useful indicator to identify fake news • Wiener Index (1): The Wiener Index of a network is
spreaders [
features 5 − 6 are stronger in close proximity to the source nodes, which is consistent with the observation in the node-level features. Hence, it further signifies the ability of the proposed label propagation scheme to preserve the network-level information.
After the local and global representations are obtained, we concatenate them to create the final embedding of the propagation network: ( ) = ( )⊕ ( ), where ⊕ is the concatenation operation.
Dataset Attribute 1 2 3 4 0.6 0.4 0.2
0 −0.2 (Δ , ) • Network depth (3): This measure captures how far the information is propagated via tweets and retweets. • Maximum outdegree (4): This characterizes the most influential node in a propagation network. • Number of nodes at diferent hops ( 5): This measure counts the number of -hop neighbours with respect to the source node in a propagation network. • Branching factor at diferent levels ( 6): For a given level in a propagation network with respect to the source node, the branching factor at is calculated as the ratio of the nodes at + 1 and the nodes at . = ||( ( (Δ ))) − ( ) ||2 (1)
Several observations can be made from the correlation analysis of 1 − 6 in Table 1 and Figure 5: (1) the where is the encoder: (Δ ) → ∈ R′ , global embeddings extracted from the initial propagation is the decoder: → ( ) ∈ R , ′ is the latent network do not show an obvious correlation with the news label; (2) the global embeddings from the complete dimension of the autoencoder, and ( ) is the prepropagation network, however, show much stronger cor- dicted embedding for the complete propagation network. relation, which demonstrates the importance of having Both () and () mappings are modelled as 2-layer access to the complete network; and (3) there is a moder- feedforward neural networks followed by a Sigmoid acate correlation between the global embeddings generated tivation function (), which can be formally defined as from the initial network and from the corresponding com- follows: plete network, which indicates the feasibility of using the initial network to predict the future embedding.
As shown in Section 3, the embedding of the complete propagation network of news records have a relatively strong correlation with the labels. However, only the initial part of the propagation network is available at the early detection deadlines. Hence, we propose to train an autoencoder that takes the partial propagation network as input, and generates the embedding of the complete propagation network.
Formally, for a given news record in the training set, denote the embedding of the initial network (
Δ ) and the complete network ( ) as (Δ ) ⊕ (Δ ) and ( ) ⊕ ( ), respectively. The autoencoder is trained using the following reconstruction loss.
( ( )) = ( 2 ( ( 1) ( ) + 1) + 2) (2) ( ) = ( 4 ( ( 3) + 3) + 4) where { 1, 4 } ∈ R(2′,) , { 2,
3 } ∈ R(′,2′) , {1, 4} ∈ R2′ , and {2, 3} ∈ R′ are trainable parameters.
We leave the optimal neural architecture search for () and () in our model as future work.
Subsequently, the generated embedding of the complete network is used to classify the news record.
= ( ( ∗ ( ) + ), ) where () is the standard binary cross entropy loss function and , are the trainable parameters of the fake news classifier.
The final loss function jointly optimises and : (3) (4) (5) = + 5. Experimental Verification
In this section, we present our experimental results to the GossipCop dataset, outperforming the best baseline demonstrate the eficiency of the proposed algorithm. by as much as 10% in accuracy.
For the PolitiFact dataset, the proposed approach
out5.1. Experimental Setup performs all the baselines except dEFEND. However, the
result for dEFEND is obtained using the complete
propaBaselines. We compare our approach with four widely gation network for each news item, while our approach
used text-based methods: (1) RST [15], (2) LIWC5, (3) only requires the initial propagation network at the
detext-CNN [
Parameter Settings. The proposed approach has erators to bypass detection—similar to the well-known
three model-specific parameters: (1) the latent embed- adversarial attacks against machine learning models.
ding dimension ′—the default value is set to 10, as we Ablation Study. Table 2 shows that without the
rehave empirically observed that the performance of the construction loss proposed in Eq. 1, i.e., the model makes
model plateaus for ′ ≥ 10; (2) detection deadline Δ—the classification only based on the embedding of the
inidefault value is set to 5 ℎ , and we also analyse the tial propagation network, which is less informative as
model performance under other values of Δ; and (3) is shown in Section 3, its accuracy drops by around 3% for
set to 0.5. As for the baselines, please refer to [
To evaluate the performance of the proposed approach, types of features. It can be seen that Node-level
Userwe adopt the commonly used metrics: (1) Accuracy (Acc); based Features are the most important among all
Node(2) Precision (Prec); (3) Recall (Rec); and (4) F1 Score level Features, which is due to the high correlation of
(F1). Following the previous works [
users. These engagements can be transferred into
different types of graphs to facilitate fake news detection.
For example, a range of models have been applied to
study the propagation patterns, including Propagation
Tree Kernel [
In addition to the above two categories, a number of
methods use a mixed strategy and rely on both news
content and associated user inter-actions over social media
to detect fake news [
This type of method uses news headlines and body con- In this work, we have designed a novel representation
tent to detect fake news. The content here is not limited learning framework for fake news early detection, by
emto text-based, but can also include visual information. bedding news propagation networks using both
globalFor example, Wang et al. [