1. Introduction
In the digital age, the proliferation of social network services (SNSs) has fundamentally transformed how information is shared and consumed. SNSs offer unprecedented opportunities for information dissemination; however, they have also become fertile grounds for the spread of fake news, thereby posing significant social and economic risks. For example, the false claim that Hillary Clinton led a human trafficking ring resulted in a shooting incident at the alleged headquarters of the organization, namely, a pizzeria (
https://www.nytimes.com/2016/12/05/business/media/comet-ping-pong-pizza-shooting-fake-news-consequences.html (accessed on 21 March 2024)). Similarly, false information suggesting Barack Obama had been injured in an explosion led to a significant economic loss, wiping out USD 130 billion in stock market value (
https://www.forbes.com/sites/kenrapoza/2017/02/26/can-fake-news-impact-the-stock-market/?sh=5e7bd2b22fac (accessed on 21 March 2024)). More recently, the spread of misinformation about COVID-19 has exacerbated societal divisions, such as those between vaccine proponents and opponents [
1]. For example, it has been reported that Republican Party members in the United States tend to support misinformation about anti-vaccination more than Democratic Party members [
2]. Fake news has been found to spread faster than factual news [
3]. Therefore, it is difficult to effectively curb its spread through manual fact checking alone to prevent such harm and societal division.
Early fake news detection strategies primarily analyzed news content to automate fact checking, whereas subsequent approaches have exploited the unique propagation patterns of news on social networks by using network graphs to represent the spread of news articles. The basis of these approaches is to represent article propagation on SNSs as a network graph, where posts are nodes and sharing relationships are edges. Using graph neural networks (GNNs) to learn this graph structure, the approach goes beyond simply examining the content of articles to identify unique features of fake news spread in the propagation process, thereby improving the detection capabilities. Additionally, previously, we proposed a fake news detection method that considers chains of similar stances or interaction patterns [
4]. This method takes advantage of confirmation bias, in which individuals perceive opinions that align with their beliefs as more persuasive, and the psychological tendency toward homophily [
5]. The method is based on studies in which researchers examined how information spreads through interactions between users with similar views [
6]. Note that the term “opinion” in this paper means “an idea or view formed about something” and is sometimes paraphrased as “stance” or a similar term. By tracking the formation of propagation between users with similar opinions while accounting for psychological characteristics, we successfully captured the complex structure of fake news spread, which led to improved detection accuracy.
Propagation-based methods offer high accuracy in detecting fake news but lack interpretability in their inferences. These models can identify fake news, but they do not provide clear, human-understandable justifications based on the analysis of propagation structures. The visualization techniques of Lu et al. [
7] and Ni et al. [
8] highlight key users and attributes in discrimination through attention weight visualization from models with a transformer-based backbone network. The efforts by Jin et al. [
9] and Yang et al. [
10] to incorporate topic, user, and network information into knowledge graphs provide external knowledge for discrimination. However, these approaches primarily visualize information emphasized during model training. They fall short of providing a consistent explanation for judging news as fake or real, in particular, a logical and systematic explanation of why the propagation structures of fake and real news differ.
Because propagation-based methods focus solely on inference without providing structure-based explanations, post-inference analysis of the propagation graph is the only alternative. Understanding the geometric characteristics of the propagation structures of fake and real news is crucial, but first, a detailed analysis of the factors that shape these structures is necessary to understand why these structural differences occur. Vosoughi et al. [
3] and Zhao et al. [
11] analyzed the depth of fake news propagation structures and reached similar conclusions. They did not identify the specific factors that lead to the formation of different propagation structures between fake and real news; however, in previous studies, researchers suggested that chains of similar opinions are closely related to the formation of fake news propagation [
5,
6,
12]. Understanding these interactions has contributed to the accuracy of propagation-based detection methods [
4]. Therefore, to improve the interpretability of propagation-based detection methods from a structural perspective, it is essential to elucidate the similarity of opinions within propagation structures.
The main goal of this study was to clarify how the propagation patterns of fake and real news differ, to identify the factors behind these differences, and to better understand the methods used to detect fake news based on these patterns. In particular, we examined the alignment of opinions within propagation structures and its relevance in the creation of fake news. This work did not aim to directly reveal the structural patterns that propagation-based detectors learn by analyzing their model parameters. Rather, it visualized the differences in propagation structures between fake and real news by analyzing opinion similarities and graph metrics within these structures, ensuring that the derived observations were statistically grounded rather than subjective. As a result, we revealed structural differences between the propagation graphs of fake news and real news, and inductively identify the predictive basis for the graph structure characterized by the propagation-based detector.
Based on the research background described above, in this paper, we propose an explanation method based on the visualization of article topics and propagation structures to improve the interpretability of graph-based fake news detectors. The proposed module provides visualization information that represents structural elements of the graph (e.g., large hubs of adjacent nodes or the depth of paths) and cluster numbers that represent document topics based on text clustering. Specifically, using BERTopic [
13], which is a state-of-the-art topic classification method, we automatically identified topics from text data and analyzed how these topics influenced patterns of news propagation. The similarity of topics could contribute to the formation of specific structural patterns in the propagation of fake news. Understanding this could improve the effectiveness of propagation-based detection methods and play a crucial role in interpreting their reasoning process. Our proposed method helped to analyze the significant structural patterns in these news propagation graphs in a form that was easily understandable by humans.
To validate the effectiveness of our proposed method, we applied the visualization module to several news propagation graphs selected from a real-world dataset and conducted a case study to observe characteristic propagation paths. We then investigated the differences in propagation structures between fake and real news and formulated hypotheses regarding the discriminative patterns of graph-based detectors. By quantifying the results and performing statistical tests based on these metrics, we verified the validity of our hypotheses and demonstrated that our proposed method provides a basis for graph-based identification of fake news.
Our contributions are threefold:
We propose a visualization method that supports the interpretation of graph-based fake news detection. This method classifies posts (i.e., nodes) in the news propagation graph by topic using BERTopic, thereby revealing chains of similar opinions, similarities within the propagation structure, and their contribution to the formation of fake news.
We provide details of an evaluation method for interpretability using the proposed method. We conducted a case study using a large-scale fake news dataset constructed from Twitter (now X) post data, observing large networks of news propagation involving many users, tweets, and quote retweets. We expressed the explanations obtained from the observations in terms of metrics, and evaluated the metrics through statistical tests. The novelty of this study lay in proposing a method to visualize the factors that form different propagation structures for fake and real news, and in defining and verifying new metrics to ensure the reliability of the provided inference base.
The observations and statistical tests provided new insights into the factors that contribute to the formation of news propagation structures, thereby contributing to the analysis of structural factors in news propagation.
In this study, we did not focus on user interface development and did not directly verify whether the proposed method contributed to improving users’ ability to identify fake news and real news, which is a future research topic. However, the reliability of the proposed method was supported by statistical validation using a large real-world dataset.
This paper is organized as follows:
Section 2 provides a comprehensive and detailed review of related studies;
Section 3 proposes a visualization module using topic classification;
Section 4 validates the proposed method with a case study on the Twitter dataset;
Section 5 provides a discussion and derives new metrics based on the insights gained. Finally,
Section 6 presents the conclusions.
2. Related Work
2.1. Content-Based Detection Methods
Content-based detection methods identify fake news based on the content of articles. These methods primarily focus on analyzing the text of articles, but some also use attached images or videos [
14,
15]. Collectively, these approaches are known as content-based detection methods. Within this category, style-based detection methods pay particular attention to the emotional language characteristic of fake news [
16], using sentiment analysis to identify words or phrases that convey strong emotions [
17]. Stance analysis, which assesses the speaker’s position or judgment on a statement, can also be performed based on content [
18]. It was suggested that a stance reflects the emotional intensity of the attitude toward a news topic [
19,
20,
21]. Recent advances in content-based methods have seen the application of transformers to text analysis [
15], although they face challenges in effectively processing long texts [
22]. However, content-based methods struggle to detect fake news, which mimics the writing style of real news, based solely on the content of the news. Therefore, the incorporation of additional features, including social context, was deemed necessary to improve the detection accuracy [
23].
2.2. Propagation-Based Detection Methods
Propagation-based methods take advantage of the social context created by news sharing on SNSs, not only the news content itself. These methods distinguish fake from real news by identifying different structural patterns in their propagation [
3]. Before GNNs became popular, recurrent neural networks were used to model the sequence of news sharing [
24]; however, this approach struggled to capture structural features. GNN-based methods have since gained traction because of their ability to model these complex structures. For example, when a news story is published on Twitter, it is shared by users through posts and further disseminated through retweets. This chain forms a graph with nodes that represent news, tweets, and retweets, and edges that represent sharing interactions. Learning this graph requires unique node features, such as user attributes like follower counts and profile descriptions, to detect users spreading fake news [
7,
8]. By contrast, news detection uses text features extracted from posts and these features associated with the corresponding post nodes are fed into the GNN for structure learning.
GNNs can learn structure through neighborhood aggregation, aggregating features of neighboring nodes through edges to the focal node for structure-aware learning. The graph convolutional network (GCN) [
25] updates a node by convolving its features with those of its neighbors and represents a basic form of a GNN. The graph transformer network (GTN) [
26] uses the transformer architecture [
27] for neighborhood aggregation by focusing on important nodes based on feature similarity between the focal node and its neighbors.
Monti et al. [
28] proposed the first propagation-based method to use a GNN. It uses GCN to learn the structure of news propagation graphs and then uses global average pooling to extract the overall graph structure for detection. Recent developments include not only extracting network features but also combining content-based detection methods to highlight textual features of news content [
8,
29] and adaptively controlling the news propagation of manipulated nodes to improve the performance [
30]. Methods that consider sociologically analyzed characteristics of fake news propagation to improve structural discrimination are also effective. User-preference-aware fake news detection (UPFD) [
31] extracts user preferences for information based on the past post history. The authors used stance analysis for feature extraction and GTN to compute the similarity of opinions between neighboring users, with the psychological phenomenon of confirmation bias playing a role in these methods [
4]. Confirmation bias, where information that supports one’s own beliefs is considered more persuasive [
5], is conceptualized as “preference” in UPFD and “homophily” in our previous research [
12]. Since opinion similarity between posts improves the performance of propagation-based methods, these results suggest that GNNs may implicitly capture opinion similarity between posts. Nevertheless, GNNs face a challenge known as the over-smoothing effect, where the distinction between user features becomes blurred due to the aggregation of neighborhood information, potentially obscuring the representation of user homophily. To solve this problem, the integration of learning strategies that account for heterophily [
32]—the potential for interactions between users from different viewpoints—was identified as a promising way to improve the performance of graph learning.
2.3. Explainable Detection Methods
Some fake news detection methods aim to visualize the model’s reasoning process. Evidence-based methods, which are derived from content-based approaches, explore similarities and differences between claim statements and texts from multiple sites used as evidence for fact checking. Recent models have become popular for highlighting these similarities and differences by visualizing the intensity of attention to texts, thereby exhibiting the importance of focused sentences or words in interpreting the dubiousness of claims [
33]. However, a limitation of evidence-based methods is their reliance on the availability of clear and reliable evidence for each claim, which becomes particularly challenging for claims with ambiguous veracity.
In this study, we focused on the implementation of interpretability within propagation-based methods; however, no existing models can explain inferences from the structure of propagation networks. We present detection methods that can visualize inference over networks, which are broadly categorized into attention-based and knowledge-graph-based methods.
Attention-based methods identify the tendencies of users who spread fake news by assigning user attributes within the propagation network. For example, Propagation2Vec [
34], which is categorized as a propagation-based method, identifies key user attributes in tweet-to-retweet cascades. The graph co-attention network (GCAN) [
7] visualizes critical user behaviors by linking user interactions with content and posting order. The multi-view attention network (MVAN) [
8] uses the graph attention network [
35] to identify user attributes that contribute to detection. GCAN focuses on words or phrases that users pay attention to, whereas MVAN highlights words that are considered indicative of falsehood.
By contrast, knowledge-graph-based methods use collective intelligence that aggregates multiple pieces of information, including user attributes and frequently occurring words in articles, to visualize reasoning behavior in detail. Jin et al. proposed a graph-kernel-dependent method, which moved toward fine-grained reasoning [
9], and its improvement, reinforcement subgraph reasoning [
10]. These methods divide the knowledge graph into subgraphs and display those composed of knowledge groups that ease news discrimination. Yang et al. discussed the effectiveness of visualization in these methods, with subject evaluations based on the visualization information, calculation of accuracy, and aggregation of confidence levels during judgment using Likert scales in [
10].
However, these methods still encounter interpretability challenges. User attributes are useful for identifying users spreading fake news; however, the attributes used for learning are not consistent across methods, and the types of user attributes to focus on differ between fake and real news, which makes it difficult to provide consistent explanations for inference. In particular, in Propagation2Vec, focusing on the same user attributes for both individual-level and cascade-level attention blurs the distinction between the attention of different individuals, which makes it difficult to determine consistent explanations.
GCAN and MVAN visualize significant words through word clouds, highlighting words that contribute to fake news judgment. MVAN and knowledge-graph-based methods [
9,
10] enable referencing sentences or words that facilitate fake news judgment. However, they focus on sentences in which users explicitly judge the content as fake, such as “confirmed” or “Sorry fake news”, which prevents the direct identification of the reasons for the news being fake from the visualization information. Supervised deep learning models learn to fit data that represent correct labels, which can contribute to increased confidence in the visualized information. However, in ambiguous truth conditions, this may only reinforce biases, such as “it’s fake news because a user said so”.
In the context of improving interpretability, our goal was to provide users with the ability to understand the reasoning behind the classification of news as fake using visualized data. This capability will empower users to assess the credibility of news stories, potentially before they are widely disseminated. By fostering a culture of critical thinking in the evaluation of information and opinions, we aimed to help mitigate the spread of fake news. However, existing methods encounter obstacles in providing explanations that are consistent and reliable across cases when distinguishing between fake and real news. “Consistent explanations” refers to the provision of reasoning that remains stable and coherent across cases.
2.4. Propagation Patterns of Fake News
Currently, no methods have been explicitly designed to elucidate propagation structures, and the factors that form different propagation structures are not fully understood. Vosoughi et al. analyzed propagation networks with convergent diffusion and reported that false information spreads from one user to another, involving many users and forming deep propagation structures [
3]. Similarly, Zhao et al. analyzed networks observed in the early stages of diffusion and showed that real news propagation tends to form dense structures, as users cluster around the disseminated information, whereas fake news forms sparser networks, with connections between users becoming more pronounced [
11]. A commonality in these reports is that fake news propagation is formed by the exchange of information transmitters from one user (A) to another (B), and from B to another user (C), and so on. Vosoughi et al. described this user-to-user sharing chain as viral and associated such propagation with a viral branching process. By contrast, dense structures in which many users share the same information, similar to the way television shows or newspapers disseminate information to a large audience, are referred to as broadcast [
3]. Viral propagation suggests a link to confirmation bias, where users share opinions that align with their own beliefs. The analysis of the factors that drive these viral-like user propagation chains remains inadequate.
There are exceptions to the findings reported by Vosoughi et al. [
3]. Notably, Jang et al. reported instances in propagation graphs used to learn fake news detection methods where real news was spread more deeply and to more users than fake news [
36]. This suggests that both real and fake news can form deep pathways as user engagement increases.
In this study, we quantified and verified patterns that can distinguish fake news propagation graphs based on insights gained from the efficient visualization of complex graph structures as the number of nodes increase. In particular, the interaction between users sharing similar opinions is believed to be deeply related to the propagation of fake news, particularly at points where user sharing is facilitated. Therefore, to explain the veracity based on structure, it is necessary to analyze not only the structural differences but also their formation through user interactions and post content. For example, the visualization might reveal that fake news often spreads through tightly knit clusters of users with high opinion alignment, whereas real news tends to spread more broadly across diverse user groups. Such insights could improve the interpretability of traditional propagation-based detection methods.
4. Experiments
In the experiment, we demonstrated the effectiveness of our newly developed visualization method through an experimental setup that consisted of four distinct steps:
- Step 1.
The proposed method was applied to the news propagation graph (
Section 4.1).
- Step 2.
Using the visualization information obtained, propagation cascades on Twitter were selected for observation within the news propagation graph and the substructures of the propagation were analyzed while interpreting the text based on the visualization information (
Section 4.2).
- Step 3.
The observations were quantified as metrics and analysis was performed over the entire graph (
Section 4.3).
- Step 4.
Statistical tests were performed based on the analysis results (
Section 4.3).
We discuss the experimental results in
Section 5. Step 4 focused on applying statistical tests to validate whether the visual cues related to distribution characteristics could discriminate between the distribution of fake and real news. This step involved generating hypotheses based on the visual data and testing their statistical significance. A positive result in this step would confirm the effectiveness of the visualization method in identifying structural variances in the propagation graphs of different types of news, thus providing a reliable predictive basis for propagation-based detectors.
4.1. Data Collection and Application of the Proposed Method to the News Propagation Graph
The observed news propagation graphs consisted of large-scale data collected from news stories that were verified by the fact-checking website PolitiFact (
https://www.politifact.com/ (accessed on 20 June 2023)) and posts that were deemed to be related to them [
4]. The data included the spread of the news on Twitter for one month after its publication, from December 2020 to July 2022. Tweets related to the news were collected using the Twitter API v2 (
https://developer.twitter.com/en/docs/twitter-api (accessed on 20 June 2023)) based on search queries generated from a list of keywords based on the frequency of word occurrences in the news. To avoid selecting overly general queries, the number of keywords used to construct the query was set to four. All combinations of the four selected keywords were searched and tweets that matched multiple queries were collected without duplication. Using common words as search terms may identify tweets that are not directly related to the news. It is difficult to manually review and remove content because of the legitimacy of the content judgment and the significant effort required. Therefore, the data collected in the study may have included noisy tweets unrelated to the news. Quote retweets were collected iteratively based on the IDs of the collected tweets. Based on this collected data, news propagation graphs were constructed. Each news propagation graph required a fact-check result and a corresponding real/fake label. In this collected data, following the examples of [
33,
43], PolitiFact judgments of true and mostly true were labeled as real, whereas mostly false, false, and pants-on-fire were labeled as fake.
The proposed visualization method was applied to the constructed news propagation graphs to obtain visualization information. This step corresponded to step 1 of the experimental procedure. The parameter MinClustSize in BERTopic was set to the default value of 10 in the official implementation (
https://github.com/MaartenGr/BERTopic (accessed on 14 December 2023)) to eliminate the parameter dependency of the proposed method. Regarding the cluster numbers in the visualization information, BERTopic outputs a special cluster number −1 for outlier data that do not belong to any clusters allowed by HDBSCAN. However, because of the transformer’s problem with long texts, BERTopic may assign the cluster number −1 to the news article text. Treating the news article text, which best represents the news propagation graph as a root node, as outlier data is inappropriate. For a fair comparison in this experiment, news propagation graphs where the news article text was assigned the cluster number −1 were excluded from the analysis.
The statistics of the graph data are shown in
Table 1. The amount of the collected post data was comparable with that of the traditional representative dataset, namely, FakeNewsNet [
43]. Thus, this dataset was sufficiently large and provided a foundation for exploring the generalizability of our findings. Data collection was limited to using the API according to Twitter’s privacy policy, whereas FakeNewsNet currently has a large number of data points that cannot be collected because of the deletion of posts and users; hence, it was excluded from the experiment.
The dataset contained 5696 hubs, where a hub was defined as any non-terminal node with more than one adjacent node. The distribution of the number of adjacent nodes for these hubs is shown in the histogram in
Figure 2, where each bin has a width of approximately 10 from the minimum value. In addition,
Figure 3 plots the number of adjacent nodes for each hub in ascending order.
Figure 3 illustrates that within our collected dataset, only a few hubs exhibited explosive propagation, which was characterized by an extremely large number of adjacent nodes to a single node, as indicated by the sharp rise in the plot.
Figure 2 and
Figure 3 show that nodes with more than 400 neighbors were in the top 1%, while nodes with 30 or fewer neighbors represented 70% of the total. For the entire dataset, there were 164,641 paths of depth 2 and 17,184 paths of depth 3 or more.
4.2. Case Studies
In this section, we follow step 2 of our experimental framework, focusing on observing news propagation, understanding its structure, and correlating it with the textual content. The goal of these case studies was to show how our proposed visualization method can help identify the cascades that are essential for distinguishing fake news from real news, and to illustrate their role in making this distinction. To ensure a fair comparison between the spread of real and fake news, we standardized both to the same order of magnitude in terms of scale.
The selection of objects to observe from the news propagation graph must be based on visualization information. First, we focused on active cascades, which contained a large number of posts and had deep propagation. By extracting paths with deep cascades and observing changes in the number of quote retweets of posted nodes in the paths and the number of adjacent nodes following them, we could understand the characteristic structures of the cascades. By selecting deep paths with post nodes that had a high number of adjacent nodes, we selected cascades with a high volume of posts. Therefore, we extracted paths with a significant depth and hubs with a large number of adjacent nodes from the visualization information as selection criteria for structures to analyze the differences in propagation.
Next, we analyzed the process of structure formation while interpreting the content of posts based on the consensus of opinions. We used cluster numbers for this analysis. A cluster number represents a group of texts with similar content to a single topic. Selecting posts with the same cluster number among adjacent posts allowed us to analyze the presence or absence of interactions between the same opinions. In particular, considering that consensus (i.e., agreement) with the opinions of others is related to the propagation of fake news, the exploration of same-topic interactions among topic transitions of hub nodes that form a deep path representing a chain of inter-user sharing is useful for analyzing the relationship between the propagation of fake news and the consistency of opinions. However, posts that were assigned a cluster number of −1, which indicates that they did not belong to any cluster, indicated a low probability of content similarity between adjacent posts. Therefore, we did not consider them in the opinion similarity analysis.
To summarize, by using the proposed method, we selected observation paths based on three criteria:
- (1)
Paths with significant depth.
- (2)
Hubs with a high number of adjacent nodes.
- (3)
Adjacent nodes with the same cluster number.
These criteria served as axes for the selection of observation paths. In the following sections, we present two cases of fake news propagation (
Section 4.2.1) and real news propagation (
Section 4.2.2).
4.2.1. Case Study for Fake News Propagation
First, we observed the example of fake news propagation shown in
Figure 4. Before explaining the graph, we describe its components. The article section shows the context of the news story, with the claim that PolitiFact is addressing underlined (this was the focus of the assessment). In the post nodes, the green-encircled text represents tweets, whereas the blue-encircled text represents quote retweets. For each post, the cluster number, full text of the post, and number of adjacent nodes are listed from top to bottom. The cluster number is highlighted in red if it matches the number of an adjacent post. In the post text, user accounts are anonymized with (@) and URLs or links are anonymized with (URL). The yellow highlighting for the number of adjacent nodes corresponds to the highlighting in the table at the bottom-right of the image. This table lists the number of adjacent nodes in descending order down to the fifth node, with the maximum depth reachable by the path through that node in parentheses. In particular, if this depth is the maximum for the entire news propagation, it is marked as “deepest”. If the observed node appears in the table, the numbers in the table are highlighted.
Figure 4 shows a large cascade in the news propagation graph for the following false claim: “In the United Kingdom, ‘70-plus percent of people who now die from COVID are fully vaccinated’”, which consisted of 27,068 nodes and 26,734 paths. We selected this cascade because of its significant depth and number of nodes. In particular, there were only two types of paths with a depth of four: paths through quote retweet nodes with 8 or 61 adjacent nodes, as shown in
Figure 4. Additionally, it is possible to visually confirm that tweets and their quote retweets in this cascade belonged to the same cluster number, and to direct attention to examining the actual content of the posts to predict the continuity of opinion, particularly when the cluster numbers were not contiguous.
The tweets that responded to this fake news introduced the content and link to a weekly report from the United States CDC [
44]. The report discussed 90 to 180 days of protection against COVID-19 based on a comparison of immunity from previous infections and vaccination. However, the authors of the report did not mention that the analysis was limited to early protection, which seems to have caused controversy. The actual report, which is not shown in
Figure 4, includes a graphical abstract of its contents (
http://dx.doi.org/10.15585/mmwr.mm7044e1 (accessed on 21 March 2024)), with annotations such as “COVID-19-like illness hospitalizations 90–179 days after previous infection or full vaccination” and “Received two doses of an RNA vaccine and no previous infection” written in smaller type than the surrounding text. A review of the post on social media shows that an image was attached to the text of the post, which made it difficult to understand the premise that it was limited to early prevention without reading the smaller text.
The graphic abstract highlights the remainder of the statement: “A study of hospitalized patients with symptoms similar to COVID-19 found... Unvaccinated people with a previous infection were 5 times more likely to have a positive COVID-19 test than vaccinated people. Get vaccinated as soon as possible”. Regardless of this, the post alone was insufficient to provide accurate information and may have led to misleading interpretations (
https://firstdraftnews.org/articles/fake-news-complicated/ (accessed on 21 March 2024)) about vaccine efficacy. Moreover, the tweet was followed by retweets of the quote that prioritized pro-vaccine claims without addressing the misconception, thereby sparking intense debate.
One path branching off from this quote retweet (the left path in
Figure 4) consisted of posts questioning the validity of the experiment, implying skepticism about the research subjects and analytical methods used to support the claim. The other path (the right path in
Figure 4) contained posts criticizing those who agreed with the report. The “Rochelle Walensky” mentioned in the text was the former director of the CDC. The last post in the path appeared to be supportive of the CDC; however, the context “In other words” was interpreted as sarcasm, implying “job well” for the CDC for releasing a report that undermined public trust, and categorizing it as critical of the CDC.
Based on the analysis, we found that fake posts typically had fewer nodes directly connected to the original post, but many deep paths extended from it. For example, while the maximum number of nodes directly connected to a post was about 120, the number of paths that extended three levels or more deep was significant. Specifically, there were more than 150 paths that extended from nodes with about 90 connections, including 60 from criticism of a former CDC director, indicating viral-like propagation. Notably, the number of paths of depth 3 or more in this subset accounted for 0.8729% of the 17,184 paths of depth 3 or more in the entire dataset. In contrast, the green-encircled tweet had the largest number of adjacent nodes in this cascade, and the number of depth 2 paths that consisted of this node was 0.07289% of the 164,641 depth 2 paths in the entire dataset. Therefore, the ratio of the number of paths of depth 3 or more to the size of the number of adjacent nodes was , indicating the predominance of viral features, i.e., mainly paths of depth 3 or more. Furthermore, this cascade showed agreement between adjacent posts, as evidenced by the clustering of posts with the same topic number. This clustering suggests that the repetition of certain phrases or sentences from the original post to emphasize its message led to a similarity in expression, which, in turn, was captured as agreement in the propagation structure of fake news. This observation underscored the role of agreement in shaping the spread of fake news.
4.2.2. Case Study for Real News Propagation
Next, we observed an example of real news propagation on a scale similar to that shown in
Figure 4. The news in the article section of
Figure 5 is an official statement made by Joe Biden, the 46th President of the United States, from the White House. The gun control law mentioned in the article text refers to the so-called Brady Bill, which was enacted in 1994 under the administration of William Jefferson Clinton, the 42nd President of the United States. This law, which was initially a five-year temporary law, was extended for another five years before expiring in 2004 under the administration of George W. Bush, the 43rd president. The news propagation graph associated with this claim consisted of 23,153 nodes and 22,936 paths.
The findings on fake news propagation presented in
Section 4.2.1 suggest that real news propagation may exhibit contrasting characteristics, such as a higher number of adjacent nodes; a reduced frequency of depth 3 paths, indicating user-to-user sharing; and an absence of consensus interactions within these paths. This section presents identified instances of propagation that exhibited these characteristics.
Figure 5 shows a cascade that satisfied the assumption of real news propagation characteristics and was discovered by using the proposed method. The green-encircled tweet following the article in
Figure 5 shows a significant hub compared with
Figure 4 and contained paths with a depth of 3, which indicates chains of user sharing. However, there were no consecutive identical cluster numbers. This tweet directly addressed the content of the claim, and notably, the poster was President Biden himself. The tweet rephrased the claim in the underlined section of the article as “The gun control law we Democrats passed was repealed by Republicans, and since then, the number of shootings has tripled”. The top two posts with the highest number of adjacent nodes among the quote retweets following this tweet are shown in
Figure 5. In the cascade on the left side of
Figure 5, there was a quote retweet that showed a sympathetic attitude toward the tweet. However, the quote retweet appeared to refer to a new gun control bill that the posting user (a Democrat Party member) claimed to have introduced, and therefore, the content was not directly identical to the target news, and the cluster numbers did not match. The path ended with a quote retweet that expressed a negative view of the passage of the gun control bill by referring to “Americans’ inherent rights”, which contradicted the values of the gun control bill and implied the Second Amendment to the United States Constitution. A quote retweet was observed that indicates that the argument regarding the effectiveness of the gun control bill, i.e., the contribution of the gun control bill to the reduction of shootings, could not be inferred, and was located at the node with 16 adjacent nodes in
Figure 5. An investigation of the authenticity of this statement found that the referenced “An Updated Assessment of the Federal Assault Weapons Ban” [
45,
46] existed. This official document published by the Office of Justice Programs of the United States Department of Justice stated that the law exempted millions of assault weapons and large-capacity magazines from regulation and that it is too early to conclude that the law contributed to a reduction in gun violence. This path ended with a post that distrusted the parent post.
As shown in
Figure 5, we observed that there existed propagation paths with a depth of 3, even in real news propagation. The number of nodes adjacent to the tweet node reached 414, which was the number of adjacent nodes that belonged to the top 1%, as shown in
Figure 2. The path propagated from this tweet node reached a depth of 3 via the quote retweet nodes with 28 or 16 adjacent nodes. The number of adjacent nodes to these quote retweet nodes was relatively small compared with the number of adjacent nodes that the tweet node had, suggesting that this was an exceptional path. Further examination shows that if we estimated the total number of paths with depth 3 or more to be 44, based on the sum of the neighboring nodes for the quote retweet nodes, this represented 0.2561% (=
) of the total dataset’s paths with depth 3 or more. On the other hand, if we estimated the total number of depth 2 paths to be 414, following the tweet node’s neighbors, this was 0.2515% (=
) of the total dataset’s depth 2 paths. Therefore, the ratio of paths of depth 3 or more to the size of the adjacent nodes was 1.018 (=
), which was about 1/10 of the ratio calculated in
Section 4.2.1 for the case of fake news. This indicates the dominant broadcast-like characteristic, where a large number of nodes cluster around widely disseminated tweets.
Furthermore, we observed that users often made their own claims and engaged in discussions in quote retweets that were not included in the original tweets, suggesting that paths with a depth greater than 3 did not necessarily share the same stance.
Figure 4 shows cases where the topic number matched between adjacent nodes due to the repetition of the original tweet’s claim. In contrast,
Figure 5 demonstrates that posts with unique claims or perspectives from users in quote retweets typically did not cluster in the BERTopic analysis, thus receiving the exceptional cluster number −1.
4.3. Hypothesis and Statistical Test
Hypothesis—Summary of the hypotheses based on the observed propagation characteristics of both real and fake news:
To extend these hypotheses to a structural analysis of all graphs in the dataset, we converted the observations into quantitative metrics, corresponding to step 3 of the experimental procedure. According to the hypotheses, real news propagation showed a broadcast structure with a large number of adjacent nodes and fewer paths with depth greater than 3, whereas fake news propagation showed a viral structure with fewer adjacent nodes, but many paths with depth greater than 3. Furthermore, the confirmation of opinion similarity within the viral structure was crucial. When observing the visualization results, we focused on paths where post nodes with the same cluster number were consecutive. Therefore, to quantify the characteristics of fake news propagation, it was necessary to count the number of paths, i.e.,
, that were adjacent between posts with the same cluster number and had a depth greater than 3. The greater the number of such paths, the more likely it was that the posts along those paths shared similar opinions. The total number of paths
in the news propagation graph scaled with the number
. Observations about the number of adjacent nodes focused on the density of a single post, and thus, could be compared with the maximum value
. Considering the lower
in inferring the spread of fake news, the observations were ultimately represented by the following intuitive metric:
where
is the total number of nodes in the news propagation graph, which was used to scale the number of adjacent nodes
. This metric compared the ratio of the maximum number of adjacent nodes to the number of paths with a depth greater than 3 that had adjacent nodes with the same cluster number, with larger values indicating fake news.
Statistical test—We expected this metric to produce larger values for fake news propagation than real news propagation. We analyzed this assumption through statistical tests that corresponded to step 4 of the experimental procedure. The distribution of the metric calculated for each news propagation graph was not normal (Shapiro–Wilk test: ); therefore, we applied the U-test to the hypothesis that “the metric values for fake news propagation do not tend to be larger than those for real news propagation”. The result was , which allowed for rejecting the hypothesis and indicates that the metric values for fake news propagation tended to be larger than those for real news propagation. This suggests that the differences in propagation characteristics outlined in the hypotheses contributed to a structure-based explanation for distinguishing fake news from real news. Specifically, this result shows that the propagation characteristics identified in the case studies, such as the viral characteristics and consistency of opinions in fake news cascades, may be representative of broader patterns in the dataset. Furthermore, this result demonstrates the effectiveness of the proposed method in supporting an efficient search for observation targets and providing visualization information for analyzing the relationship between the graph structure and post content.
6. Conclusions
In this study, we aimed to elucidate the differences in the dissemination structures between fake news and real news, in addition to the specific factors that contributed to these structures, thereby enhancing the interpretability of propagation-based fake news detection methods. To achieve this, we proposed a visualization module that illustrates various aspects of news propagation, such as the number of adjacent nodes, the depth of propagation paths, the distribution of topics, and the sequence of topics within a propagation cascade. The initial segment of our module facilitates the identification of significant cascades, especially when analyzing large volumes of posts related to specific news, highlighting the structural differences in the spread of fake versus real news. Uniquely, the latter part of our module allows for the examination of opinion similarity within these cascades, including the level of agreement with the news content, which is a feature that was not present in previous research.
The main findings from this study were twofold:
- (1)
Our experiments, which used the proposed visualization method, first showed that fake news tended to exhibit viral-like propagation patterns, whereas real news was more likely to spread across a wider network in a broadcast-like manner. A key factor in these propagation patterns was the similarity of opinions within the structures. Specifically, the spread of fake news was concentrated among user groups with high opinion similarity, contributing to its viral spread. By contrast, real news was disseminated by a diverse range of users, thereby showing broadcast-like characteristics.
- (2)
Furthermore, our analysis suggests that the propagation of fake news might have been more prevalent within specific echo chambers, indicating that such news is often circulated within insular groups of like-minded individuals. Conversely, real news appears to be more widely accepted across a broader audience. These insights underscore the significance of considering opinion alignment when analyzing the propagation mechanisms of fake and real news.
We observed the insights gained through visualization in this study in case studies and validated them for the entire dataset through statistical tests.
However, this study had limitations. The analysis of propagation structures was based on statistical features and similarities, and did not fully account for the social, cultural, and psychological factors that could influence individual cases. Specifically, the text analysis using BERTopic, which is used in the proposed method, is not able to capture these factors. These factors can play a significant role in the spread of fake news and remain an area for future research.
The findings of this study contribute to the early identification of fake news and the development of prevention measures, in addition to enhancing the credibility of real news. Understanding the dissemination structures could provide new approaches for detecting and addressing fake news characteristics. For social media platforms and users, this research offers valuable insights for assessing information quality and preventing the spread of misinformation in digital environments.
Future research should address the limitations of this study by diversifying datasets, integrating social and cultural factors, and conducting more detailed analyses of dissemination structures. Understanding the psychological mechanisms involved in fake news propagation and changes in user behavior patterns also remains a crucial area of research.