This paper introduces a novel rumor detection model for social media that enhances identification accuracy by incorporating user behavior alongside traditional user features. Utilizing graph convolutional networks for user representation, a recurrent neural network for analyzing propagation tree structures, and an integrator for merging these analyses, the model adeptly captures both user behaviors and the dynamics of rumor spread. Tested on Twitter15 and Twitter16 datasets, it achieved superior accuracy rates of 85.2% and 87.3%, respectively, outperforming existing models. Although the model currently does not differentiate interaction stances between users through weighted graph edges, its integration of user behavior marks a significant advancement in precise rumor detection.
The proliferation of rumors on social media,
highlighted by their significant impact during events
such as the 2016 U.S. presidential election [
The methodology centers around enhancing rumor detection in social media through the construction of a dataset composed of tuples representing declarations (tweets) and their associated users, arranged to form propagation trees and a user cooccurrence graph. The aim is to classify these tuples into categories of rumors (non-rumor, false rumor, true rumor, unconfirmed rumor) using a novel detection model.
Utilizes graph convolutional networks (GCN) to encode user behavior and static characteristics into a higher-order user representation. This involves processing an undirected graph comprising users linked based on their interactions, with adjacency matrix adjustments reflecting the significance of these interactions. The user features are represented as a matrix, and the GCN updates node feature matrices to integrate information from neighbor nodes.:
Employs bottom-up and top-down recurrent neural network encoders to capture the structural and semantic features of rumor propagation trees. The bottom-up approach aggregates child node representations to compute a parent node's representation, capturing long-distance interaction dependencies. Conversely, the top-down approach considers the current node's features and its parent node representation. The aim is to encode the propagation tree's structure and semantics into a vector representation.
Integrator: Combines the output of the user encoder and propagation tree encoder. It fuses the user representation with the propagation tree representation through a fully connected layer, aiming to accurately predict the category of each information statement by considering both user and propagation tree information.
Focuses on minimizing the cross-entropy loss between the model's predicted probability distribution and the true labels, incorporating a regularization term to balance the loss and prevent overfitting.
The study conducts experiments using Twitter datasets (Twitter15 and Twitter16), comprising 1381 and 1181 propagation trees respectively, and involving hundreds of thousands of users. These datasets are categorized into non-rumor, false rumor, true rumor, and unknown rumor, and split in a 9:0.5:0.5 ratio for training, validation, and test sets. Model performance is evaluated based on overall accuracy and class-specific F_1 scores.
Implemented in Pytorch, the model employs recurrent neural networks and graph convolutional networks, with evaluation through 5-fold cross validation. Key configuration parameters include a 256-dimensional word vector, 256-dimensional hidden layers for user statistical features, behavioral information, and the integrator module, and a 256sized batch with a 0.005 learning rate using the Adam optimizer. Experiments leverage Python 3.7 on a system with an NVIDIA Geforce RTX 2080 GPU. User age data undergoes preprocessing to remove unrealistic values.
The experimental analysis validates the IRUGCN
model's performance in rumor detection on Twitter
datasets, comparing it with existing methods like
BERT[
This highlights the importance of integrating user behavioral data for accurate rumor identification.
Comparative tests reveal that direct use of user statistical features or fully connected layers for user behavior analysis is less effective than the proposed user encoder, especially on larger datasets, underscoring the encoder's efficiency in capturing complex user interactions.
From the results in Table 2, we can see that our Transformer have relatively low accuracy because model outperforms other models on these 2 datasets, they do not integrate many user features and other especially TD-IRUGCN is more superior than BU- propagation features.
IRUGCN.
Compared with DDGCN and UMLARD methods, 3.2.2. Encoder Effectiveness Analysis the model in this paper considers not only the Further analysis of the user encoder's effectiveness statistical characteristics of users but also their highlighted its superiority over benchmark methods behavioral information. Since the group behavior of relying solely on user's statistical features or using users is more likely to spread rumors, the accuracy of fully connected layers for feature integration. The rumor detection can be significantly improved when study revealed that the user encoder performs user behavioral information is taken into account, exceptionally well on larger datasets, benefiting from although it may also bring about other rumor rich user behavior and interaction patterns. detection interferences. In contrast, BERT and
Table 2 Performance of bottom-up propagation tree encoder combining different features on Twitter15 dataset
Twitter15
Non-rumor Fake rumor True Rumor
A series of ablation experiments were conducted to dissect the contribution of various model components. These experiments underscored the significance of user behavioral features and revealed a comparative advantage of the top-down propagation tree structure encoder over the bottom-up approach, attributing this to its early incorporation of global information. An early rumor detection analysis further confirmed IRUGCN's effectiveness, with the model outperforming counterparts at critical early detection time points, showcasing its potential in curbing rumor spread proactively.
The study also includes a qualitative analysis of false rumor propagation patterns through sample tree cases, illustrating the dynamic interplay of support, rebuttal, and skepticism in rumor spread. This underlines the model's ability to capture complex propagation dynamics, contrasting with traditional models' limitations in grasping the depth of user interactions and responses.
This chapter outlines the development and validation of an innovative graph neural network-based rumor detection model that significantly enhances accuracy and real-time performance by incorporating user behavior. The model integrates three pivotal components: a user encoder, a propagation tree structure encoder, and an integrator, facilitating a multi-dimensional analysis of rumors through content, user, and propagation studies.
Empirical tests on real-world datasets have underscored the model's superiority in rumor detection accuracy over existing methodologies, laying a robust groundwork for future explorations. Moving forward, the research will delve into a more nuanced examination of inter-user interactions, such as assigning weights to user graph edges based on user interaction stances, to refine the model's sensitivity towards intricate user relationships. Additionally, the potential of transforming datasets into graph structures is being considered to further elevate the model's performance and versatility.
This paper is supported by the National Natural Science Foundation of China under contract No. 72074108, Special Project of Nanjing University Liberal Arts Youth Interdisciplinary Team (010814370113), Jiangsu Young Social Science Talents, and Tang Scholar of Nanjing University.