Exploiting Tri-Relationship for Fake News Detection

Social media for news consumption is becoming popular nowadays. The low cost, easy access and rapid information dissemination of social media bring benefits for people to seek out news timely. However, it also causes the widespread of fake news, i.e., low-quality news pieces that are intentionally fabricated. The fake news brings about several negative effects on individual consumers, news ecosystem, and even society trust. Previous fake news detection methods mainly focus on news contents for deception classification or claim fact-checking. Recent Social and Psychology studies show potential importance to utilize social media data: 1) Confirmation bias effect reveals that consumers prefer to believe information that confirms their existing stances; 2) Echo chamber effect suggests that people tend to follow likeminded users and form segregated communities on social media. Even though users social engagements towards news on social media provide abundant auxiliary information for better detecting fake news, but existing work exploiting social engagements is rather limited. In this paper, we explore the correlations of publisher bias, news stance, and relevant user engagements simultaneously, and propose a Tri-Relationship Fake News detection framework (TriFN). We also provide two comprehensive real-world fake news datasets to facilitate fake news research. Experiments on these datasets demonstrate the effectiveness of the proposed approach. Introduction People nowadays tend to seek out and consume news from social media rather than traditional news organizations. For example, 62% of U.S. adults get news on social media in 2016, while in 2012, only 49 percent reported seeing news on social media. However, social media for news consumption is a double-edged sword. The quality of news on social media is much lower than traditional news organizations. Large volumes of “fake news”, i.e., those news articles with intentionally false information, are produced online for a variety of purposes, such as financial and political gain (Klein and Wueller 2017; Allcott and Gentzkow 2017). Fake news can have detrimental effects on individuals and the society. First, people may be misled by fake Copyright c © 2018, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. http://www.journalism.org/2016/05/26/news-use-acrosssocial-media-platforms-2016/ news and accept false beliefs (Nyhan and Reifler 2010; Paul and Matthews 2016). Second, fake news could change the way people respond to true news. Third, the widespread of fake news could break the trustworthiness of entire news ecosystem. Thus, it is important to detect fake news on social media. Fake news is intentionally written to mislead consumers, which makes it nontrivial to detect simply based on news content. Thus, it is necessary to explore auxiliary information to improve detection. For example, several style-based approaches try to capture the deceptive manipulators originated from the particular writing style of fake news (Rubin and Lukoianova 2015; Potthast et al. 2017). In addition, previous approaches try to aggregate users’ responses from relevant social engagements to infer the veracity of original news (Castillo, Mendoza, and Poblete 2011; Gupta, Zhao, and Han 2012). The news ecosystem on social media involves three basic entities, i.e., news publisher, news and social media users. Figure 1 gives an illustration of such ecosystem. In Figure 1, p1, p2 and p3 are news publishers who publish news a1, . . . , a4 and u1, . . . , u6 are users who have engaged in posting these news. In addition, users with similar interests can also form social links. The tri-relationship among publisher, news, and social engagements contains additional information to help detect fake news. First, sociallogical studies on journalism have theorized the correlation between the partisan bias of publisher and news contents veracity (Gentzkow, Shapiro, and Stone 2014; Entman 2007), where partisan means the perceived bias of the publisher in the selection of how news is reported and covered. For example, in Figure 1, for p1 with extreme left partisan bias and p2 with extreme right partisan bias, to support their own partisan, they have high degree to report fake news, such as a1 and a3; while for a mainstream publisher p3 that has least partisan bias, she has lower degree to manipulate original news events, and is more likely to write true news a4. Thus, exploiting publisher partisan information can bring additional benefits to predict fake news. Second, mining user engagements on social media towards the news also help fake news detection. Different users have different credibility levels on social media, and https://www.nytimes.com/2016/11/28/opinion/fake-newsand-the-internet-shell-game.html? Figure 1: Tri-relationship among publishers, news pieces, and social media users for news dissemination ecosystem. the credibility score which means “the quality of being trustworthy” (Abbasi and Liu 2013) has a strong indication of whether the user is more likely to engage fake news or not. Those less credible users, such as malicious accounts or normal users who are vulnerable to fake news, are more likely to spread fake news. For example, u2 and u4 are users with low credibility scores, and they tend to spread fake news more than other highly credible users. In addition, users tend to form relationships with like-minded people. For example, user u5 and u6 are friends on social media, so they tend to engage those news that confirm their own views, such as a4. Publisher partisan information can bridge the publishernews relationship, while social engagements can capture the news-user relationship. In other words, they provide complementary information that has potential to improve fake news prediction. Thus, it’s important to integrate these two components and model the tri-relationship simultaneously. In this paper, we study the novel problem of exploiting trirelationship for fake news detection. In essence, we need to address the following challenges (1) how to mathematically model the tri-relationship to extract news feature representations; and (2) how to take the advantage of tri-relationship learning for fake news detection. In an attempt to address these challenges, we propose a novel framework TriFN that captures the Tri-relationship for Fake News detection. The main contributions are: • We provide a principled way to model tri-relationship among publisher, news, and relevant user engagements simultaneously; • We propose a novel framework TriFN that exploits trirelationship for fake news prediction; and • We evaluate the effectiveness of the proposed framework for fake news detection through extensive experiments on newly collected real-world datasets. Problem Statement Even though fake news has been existed for long time, there is no agreed definition. In this paper, we follow the definition of fake news that is widely used in recent research (Shu et al. 2017; Zubiaga et al. 2017; Allcott and Gentzkow 2017), which has been shown to be able to 1) provide theoretical and practical values for fake news topic; and 2) eliminate the ambiguities between fake news and related concepts. DEFINITION 1 (FAKE NEWS) Fake news is a news article that is intentionally and verifiably false. Let A = {a1, a2, ..., an} be the set of n news articles and U = {u1, u2, ..., um} be the set of m users on social media engaging the news spreading process. We denote X ∈ R as the news feature matrix. Users can become friends with other users and we use A ∈ {0, 1} to denote the user-user adjacency matrix. On social media sites, users can easily share, comment and discuss about the news pieces. This kind of social engagements provide auxiliary information for fake news detection. We denote the social news engagement matrix as W ∈ {0, 1}, where Wij = 1 indicate that user ui has engaged in the spreading process of the news piece aj ; otherwiseWij = 0. It’s worth mentioning that we focus on those engagements that show that users agree with the news. For example, For example, we only utilize those users that directly post the news, or repost the news without adding comments. More details will introduced in Section . We also denote P = {p1, p2, ..., pl} as the set of l news publishers. In addition, we denote B ∈ R as the publisher-news relation matrix, and Bkj = 1 means news publisher pk publishes the news article aj ; otherwise Bkj = 0. We assume that the partisan labels of some publishers are given and available. We define o ∈ {−1, 0, 1} as the partisan label vectors, where -1, 0, 1 represents left-, neutral-, and right-partisan bias. Similar to previous research (Shu et al. 2017), we treat fake news detection problem as a binary classification problem. In other words, each news piece can be true or fake, and we use y = {y1;y2; ...;yn} ∈ R n×1 to represent the labels, and yj = 1 means news piece aj is fake news; yj = −1 means true news. With the notations given above, the problem is formally defined as, Given news article feature matrix X, user adjacency matrix A, user social engagement matrix W, publishernews publishing matrix B, publisher partisan label vector o, and partial labeled news vector yL, we aim to predict remaining unlabeled news label vector yU . A Tri-Relationship Embedding Framework In this section, we propose a semi-supervised detection framework by exploiting tri-relationship. The idea of modeling tri-relationship is demonstrated in Figure 2. Specifically, we first introduce the news latent feature embedding from news content, and then show how to model user social engagements and publisher partisan separately; At last, we integrate the components to model tri-relationship and provide a semi-supervised detection framework. A Basic Model for News Content Embedding The inherent manipulators of fake news can be reflected in the news content. Thus, it’s important to extract basic feature representation from news text. Recently, it has been Figure 2: The tri-relationship embedding framework. shown that nonnegative matrix factorization (NMF) algorithms are very practical and popular to learn document representations (Xu, Liu, and Gong 2003; Shahnaz et al. 2006; Pauca et al. 2004) . It tries to project the document-wordmatrix to a joint latent semantic factor space with low dimensionality, such that the document-word relations are modeled as inner product in the space. Specifically, giving the news-word matrixX ∈ R + , NMF methods try to find two nonnegative matricesD ∈ R + and V ∈ R t×d + by solving the following optimization problem, min D,V≥0 ‖X−DV ‖F + λ(‖D‖ 2 F + ‖V‖ 2 F ) (1) where d is the dimension of the latent topic space. In addition, D and V are the nonnegative matrices indicating low-dimensional representations of news and words. Note that we denote D = [DL;DU ], where DL ∈ R r×d is the news latent feature matrix for labeled news; while DU ∈ R (n−r)×d is the news latent feature matrix for unlabeled news. The term λ(‖D‖F + ‖V‖ 2 F ) is introduced to avoid over-fitting. With the basic model for news latent representation, next we introduce our solution to model i) the relationship between news and user social engagements, and ii) the relationship between news and publisher partisans. News-User Social Engagements Embedding The social engagements of users towards news articles have added value to guide the learning process of news latent features. Specifically, as shown in the yellow block in Figure 2, we explore i) user-user relations that are used to learn the basic user latent features; and ii) user-news engagement relations that encoding the correlations between user credibilities and news features guided by news veracity labels. Basic User Feature Representation. On social media, people tend to form relationship with like-minded friends, rather than those users who have opposing preferences and interests. Thus, users that are connected are more likely to share similar latent interests towards news pieces. We use nonnegative matrix factorization method to learn the user latent representations (Tang, Aggarwal, and Liu 2016; Wang et al. 2017). Specifically, giving user-user adjacency matrix A ∈ {0, 1}, we learn nonnegative matrix U ∈ R m×d + by solving the following optimization problem, min U,T≥0 ‖Y ⊙ (A−UTU )‖F + λ(‖U‖ 2 F + ‖T‖ 2 F ) (2) whereU is the user latent matrix,T ∈ R + is the user-user correlation matrix, and Y ∈ R controls the contribution of A. Since only positive samples are given in A, we first set Y = sign(A) and then perform negative sampling and generate the same number of unobserved links and set weights as 0. λ(‖U‖F + ‖T‖ 2 F ) is to avoid over-fitting. Capturing relations of User Engagements and News The user engagements of news on social media has potential to provide rich auxiliary information to help detection fake news. However, users can express rather different and diverse opinions towards the news when spreading it, such as agree, against, neutral. In this paper, we focus on those engagements that agree with the news which can be directly implied in user actions. For example, we only utilize those users that directly post the news, or repost the news without adding comments. Those users that have different opinions are usually unavailable and needed to be inferred. To model the user engagements, we consider the inherent relationship between the the credibilities of users and their posted/shared news pieces. Intuitively, we assume that users with low credibilities are more likely to spread fake news, while users with high credibilities are less likely to spread fake news. For example, low credibility users could be that 1) users that aim to spreading the diffusion scope of fake news; or 2) users that are susceptible to fake news. We adopt the existing method in (Abbasi and Liu 2013) to measure user credibility scores, which is one of the practical approaches. The basic idea in (Abbasi and Liu 2013) is that less credible users are more likely to coordinate with each other and form big clusters, while more credible users are likely to from small clusters. Thus, basically, the credibility scores are measured through the following major steps: 1) detect and cluster coordinate users based on user similarities; 2) weight each cluster based on the cluster size. Note that for our fake news detection task, we do not assume that credibility directly provided but infer the credibility score from widely available data, such as user-generated contents (Abbasi and Liu 2013). Each user has a credibility score and we use c = {c1, c2, ..., cm} to denote the credibility score vector, where a larger ci ∈ [0, 1] indicates that user ui has a higher credibility. Since the latent features for low-credibility users are close to fake news latent features, while those of highcredibility users are close to true news latent features, we solve the following optimize problem,