Deep reinforcement learning based ensemble model for rumor tracking
Introduction
The high-speed development of social networks brings information on a massive scale. The convenience of the social networks accelerates the diffusion of information. People create and publish messages without any limitations, which leads to the widespread dissemination of rumors in social networks [1], [2], [3], [4]. Rumors usually appear with unjudged veracity and contain substantive misinformation, which may cause huge damage to the society. For instance, a large number of rumors arise during the COVID-19 epidemic, such as “More than ten thousand people die in Wuhan”. and “Liquor kills the virus”., causing great public panic. Therefore, it is an urgent task to defeat rumors in social networks.
Social media generates huge amounts of data all the time. It is impractical to manually defeat rumors on such data scalability. Therefore, many studies have been proposed and can be divided into two categories: diffusion-based rumor source identification and content-based rumor detection. Diffusion-based rumor source identification aims to locate the sources of the rumor at the early stage of rumor diffusion [5], [6], [7]. Content-based rumor detection aims to judge the veracity of the given content in social networks. As shown in Fig. 1, this task is a pipeline that consists of four sub-tasks [2], [8]: rumor detection, rumor tracking, sentence classification, and veracity classification. Rumor detection is to determine whether a tweet sequence is a rumor [9], [10], [11], [12]. Sentence classification is to judge the emotion of a tweet sentence [13], [14], [15]. Rumor veracity is to determine if a tweet tells the truth [8], [16], [17]. These three sub-tasks have attracted extensive attention in recent years. However, only a few studies have been proposed for the rumor tracking task.
Rumor tracking is to collect relevant tweets of a given rumor event. It can be considered as a binary classification problem that classifies the posts to be related or unrelated. However, the existing proposals [18], [19] only consider rumor tracking as an auxiliary task in multi-task learning without special optimization, hence restraining the accuracy of tracking performance. Therefore, we propose a deep reinforcement learning based ensemble model for rumor tracking (RL-ERT).
The main contributions of this work are summarized as follows:
- •
By exploring plenty of basic models, we propose an aggregated model named RL-ERT to solve the rumor tracking task.
- •
We analyze the rumor tracking task and find suitable features and embedding methods. Also, we propose a reinforcement learning based ensemble algorithm to aggregate basic models into a unity.
- •
We conduct experiments on public benchmark datasets, and the experimental results show the rationality and superiority of RL-ERT.
The rest of this article is structured as follows. In Section 5, we introduce important works related to RL-ERT. In Section 2, we introduce some notations and background knowledge of this work. Our proposed model RL-ERT is introduced in Section 3. Then, Section 4 shows the experimental settings and results. Finally, we give the conclusion and future work in Section 6.
Section snippets
Preliminaries
In this section, we proceed to introduce the notations, problem definitions, and background knowledge of rumor tracking. The frequently used notations are summarized in Table 1.
Deep reinforcement learning based ensemble model
RL-BRT model is inspired by Mixture of Experts model (MoE) [29]. MoE is an ensemble model that contains multiple separate sub-models. Each sub-model is an expert that is trained on a region of input data. The data distribution of each tweet is different and we believe each rumor tweet has its suitable classifier. Therefore, instead of dividing the input data into different regions, RL-ERT adopts a policy gradient based neural network to generate a weight for each sub-model and then aggregates
Experiment
In this section, we proceed to introduce the experimental settings and results. Our GPU device is Tesla P100 with 16GB memory. We use python 3.7 based Keras to implement all the experiments.
Rumor tracking
The earliest Rumor tracking work is traced back to 2011. [2] aims to find the relevant rumors for a known prior. Qazvinian et al. [18] explore three types of features including content-based, network-based, and tweet specific memes for identifying rumors. Hamidian et al. [34] devise novel features and classify rumors with WEKA platform. The previous studies mainly focus on building features with simple classifiers. Cheng et al. [19] propose a multi-task learning model named VRoC and treat rumor
Conclusion
Rumor Tracking is a valuable sub-task of automatically rumor defeating. In this paper, we propose a reinforcement learning based ensemble model named RL-ERT. We promote the performance of the rumor tracking task on accuracy and macro-F1 score with specific features and WTPN based ensemble strategy. Besides, well-designed experiments are carried out to demonstrate the effectiveness of RL-ERT. Experimental results on benchmark datasets suggest that our model outperforms the baseline methods by
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This research is supported in part by the NSFC (Grants No. 61902134, 62011530437), the Hubei Natural Science Foundation (Grant No. 2020CFB871), and the Fundamental Research Funds for the Central Universities (HUST: Grants No. 2019kfyXKJC021, 2019kfyXJJS091). Changyin Luo is supported in part by the Hubei Natural Science Foundation (Grant No. 2017CFB135).
References (41)
- et al.
Online social network analysis: A survey of research applications in computer science
(2015) - et al.
Detection and resolution of rumours in social media: A survey
ACM Comput. Surv.
(2018) - et al.
Word of mouth: Rumor dissemination in social networks
- et al.
The spread of true and false news online
Science
(2018) - et al.
Detecting sources of computer viruses in networks: theory and experiment
- et al.
Rumors in a network: Who’s the culprit?
IEEE Trans. Inf. Theory
(2011) - et al.
Finding effectors in social networks
- et al.
All-in-one: Multi-task learning for rumour verification
- et al.
Exploiting context for rumour detection in social media
- et al.
Detect rumors on Twitter by promoting information campaigns with generative adversarial learning