BiMGCL: rumor detection via bi-directional multi-level graph contrastive learning

Feature-based rumor detection

In recent years, because of the effectiveness of machine learning in handling huge numbers of labeled samples, feature-based approaches to rumor detection have been widely studied with the aim of extracting representative features to detect rumor events.

The most crucial step in traditional rumor detection methods is selecting and extracting prominent features from the data. Castillo, Mendoza & Poblete (2011) studied Twitter posts on popular topics and extracted 68 features related to user posting and retweeting, original post topics, and post text. They then used decision trees for automatic classification. Yang et al. (2012) were the first to analyze and detect rumors on Chinese microblog. They extracted two new features, posting location and posting client, from the data and combined them with information text and propagation structure features, using support vector machine (SVM) for classification. Kwon et al. (2013) investigated the time, structure, and text of rumor propagation and proposed a new cyclical time-series model. Ma et al. (2015) replaced manual feature extraction with a tree kernel method and proposed propagation tree kernel, a method extracts higher-order features that distinguish different types of rumors by comparing the similarities between propagation tree structures and then uses SVM for rumor recognition.

In recent years, many rumor detection methods have employed deep learning techniques for feature extraction and classification of rumors. Ma et al. (2016) were the first to use deep learning models for rumor detection on microblogs. They utilized the loops formed by connections between units in recurrent neural networks (RNNs) to capture dynamic temporal signal features of rumor propagation. Early rumor detection could be achieved through complex repeating units and additional hidden layers. Similar approaches include long short-term memory (LSTM) (Liu & Wu, 2018), gated recurrent units (GRU) (Lu & Li, 2020) and CNN (Yu et al., 2017). Because the temporal structure features can only represent the sequential propagation of rumors, these methods still have significant shortcomings. A large number of abundant and important structural features are also hidden by the spread of rumors, or the forwarding relationships.

However, most of the above-mentioned feature-based techniques overuse isolated features and are unable to identify the interaction between different rumor messages as well as the propagation patterns of rumors. To solve these problems, BiMGCL introduces bi-directional graph to simulate the interactions between different rumors, which can effectively capture the intrinsic interactions between rumor messages.

Graph-based rumor detection

In contrast to feature-based methods, various graph-based approaches have been put forth recently. These approaches perform better because they place a greater emphasis on the structure data of rumor events. In order to capture rich structural information for rumor detection, the global-local attention network (GLAN) (Yuan et al., 2019) models the global relationship between all source tweets, retweets, and users as a heterogeneous graph. This method produces a better integrated representation for each source tweet by co-encoding local semantic and global structural information and fusing the semantic information of relevant retweets with attention mechanisms. The Factual News Graph (FANG) (Nguyen et al., 2020) enhances the quality of the representation by capturing more social structure and engagement patterns based on the heterogeneous network. To learn the correlation between the source tweet and retweet propagation as well as the co-influence between the source tweet and user engagement, A graph-aware co-attention network (GCAN) (Lu & Li, 2020) develops a dual co-attention technique based on the graph structure. Rumor2vec (Tu et al., 2021) proposes the concept of the union graph to merge multiple propagation trees into a large graph when there are shared nodes and edges.

Graph convolutional networks (GCN) have been designed to enable neural networks to learn structured data representations. Based on CNNs, GCNs define convolution operations on graph structures, allowing nodes to integrate features with their neighborhoods. They are widely used in graph node classification and graph structure classification tasks. Bi-GCN (Bian et al., 2020) encodes and trains bi-directional rumor structures using graph convolutional networks, marking the first study to use GCNs for social media rumor detection. EBGCN (Wei et al., 2021) explore propagation uncertainty for rumor detection, which adaptively rethinks the reliability of latent relations by adopting a Bayesian approach. In recent studies, UMLARD (Chen et al., 2022) and TISN (Luo et al., 2022) employ GCN to learn the patterns of rumor propagation,they models rumor diffusion from multi-view perspective, using more information about rumors. However, they all relies on a large amount of labeled data, and supervised training struggles to cope with the real-time changes in social networks. RDEA (He et al., 2021) addresses this issue by using a self-supervised training approach, it expands data through various data augmentation methods and trains the encoding network using contrastive learning methods, ultimately achieving rumor event detection. However, similar to the issues present in the recent studies of UMLARD and TISN, they only utilizes the unidirectional propagation structure of rumor events, and thus, cannot deeply mine the dispersion structure information of rumor samples, resulting in significant limitations.

Problem definition

Assume that $C=\left\{c_1,c_2,\dots ,c_m\right\}$ is the dataset for rumor detection, where c_i represents the ith rumor event. $c_i=\left\{r_i,{\omega }_1^i,{\omega }_2^i,\dots ,{\omega }_{n_i-1}^i,G_i\right\}$, where n_i denotes the number of posts in this event, r_i is the source post, ${\omega }_j^i$ is the jth responsive post, and G_i represents the propagation structure. Specifically, we use G_i to denote a graph $〈V_i,E_i〉$, where $V_i=\left\{r_i,{\omega }_1^i,{\omega }_2^i,\dots ,{\omega }_{n_i-1}^i\right\}$ and $E_i=\left\{e_{st}^i|s,\right\}$t =0 $,\left\{\dots ,n_i-1\right\}$ denotes the set of edges between responded posts, retweeted posts, or responsive posts, and r_i is the root node (Ma, Gao & Wong, 2018). The feature matrix is denoted as $X_i=\left[x_0^i,x_1^i,\dots ,x_n^i\right]$, where $x_j^i\in {\mathbb{R}}^d$ represents the jth element of X_i and d denotes the dimension of the feature. Let $A_i\in {\left\{0,1\right\}}^{n_i\times n_i}$ be the adjacency matrix with the initial value as (1)${\displaystyle a=\left\{\begin{array}{c}\hfill 1,if{e}_{st}^i\in E_i\hfill \\ \hfill 0,otherwise\hfill \end{array}\right..}$ Additionally, our goal is to train a classifier f:C → Y, where C represents the set of events and Y denotes the fine-grained classes $\left\{N,F,T,U\right\}$ (i.e., Non-rumor, False Rumor, True Rumor, and Unverified Rumor). Furthermore, during the model training phase, we generate $\hat{G}$ through data augmentation, and subsequently a classifier f(⋅) is trained together with the original G.

Bi-directional graph construction

This article particularly focuses on global features based on the overall structure of rumour propagation direction, which is an important cue for rumour detection. The depth propagation features and breadth propagation features of rumours play an important role in rumour detection. Unfortunately, traditional temporal structure features and single rumour spread features ignore the overall structural features of rumour events, and the inability to exploit these meaningful interactions of information from different propagation directions leads to a degradation in the performance of these methods. Obviously, bi-directional graphs are a powerful tool for modelling complex interactions between information and relationships in different directions. Therefore, to settle Issue1, using bipartite graphs to model the deep spread of rumours in relational chains and the wide spread of rumours in social groups is an elegant choice for the rumour detection task.

We construct two different directions for each sample g_i: a top-down direction of propagation and a bottom-up direction of dispersion, i.e., for the direction of rumor propagation, if ${\omega }_2^i$ has a response to ${\omega }_1^i$, there will be an directed edge ${\omega }_1^i\stackrel{}{\to }{\omega }_2^i$, while for the direction of dispersion, there will be an directed edge ${\omega }_2^i\stackrel{}{\to }{\omega }_1^i$. We use $g_i^{TD}$ and $g_i^{BU}$ to represent these two cases, where the adjacency matrix $A_i^{TD}$ in $g_i^{TD}$ contains the top-down part of A_i and the adjacency matrix $A_i^{BU}$ in $g_i^{BU}$ contains the bottom-up part of A_i. Also the same feature matrix X_i is used for both directions, i.e., $X_i^{TD}=X_i^{BU}=X_i$.

Finally, as shown in Fig. 1A, we get the fine-grained bi-directional graphs of rumor samples, which are represented as adjacency matrices $A_i^{TD}$, $A_i^{BU}$, and attribute matrix X_i.

Multi-strategy graph augmentation

Recent research has shown that the performance of contrast learning is directly influenced by the samples of positive and negative instance pairs (Yin et al., 2022; Miao et al., 2022; Hassani & Khasahmadi, 2020). BiMGCL differs from existing contrastive learning methods (Li et al., 2022; Zhu et al., 2021; Qiu et al., 2020; You et al., 2020) which only deals with geometric information in graphs or pure dependencies in a single isomorphic graph, it aims to produce realistic rumor propagation augmented instances using both top-down and bottom-up types. Therefore, to resolve issue 2, we take into account both the bi-directional graph’s structure and attribute information to generate augmented positive and negative instances. By imposing structural prior or semantic prior, we focused on two directions of rumor propagation graph data augmentation methods. As illustrated in Fig. 1, responded posts attribute masking, which merely modifies the attribute matrix, is a semantic level augmentation method. On the contrary, subsampling and edge modification belong to structure-level augmentation, and they transform the adjacency matrix.

Multi-level contrastive learning-based rumor detection

To address issue 3, we use a graph encoder to learn graph-level low-dimensional representations of different rumor propagation instances. We utilize the Graph Isomorphism Network (GIN) (Xu et al., 2018), a powerful graph neural network, as the graph encoder in BiMGCL. GIN has been proven to outperform other GNNs, such as GCN (Kipf & Welling, 2016) and GraphSAGE (Hamilton, Ying & Leskovec, 2017).

After the original rumor tree samples have undergone the three types of augmentations, a rich set of sample instances has been generated. We use the GIN network to learn the hidden structures and semantic information in these tree structures, with the detailed architecture shown in Fig. 1C. In this process, we place great emphasis on the root node’s features, as the source posts of rumor events often contain rich information. By making full use of this, the neural network can learn more accurate abstract semantics of rumor propagation.

In Fig. 1C, for the ith rumor graph sample g_i = (A_i, X_i), where A_i includes both $A_i^{TD}$ and $A_i^{BU}$ (TD and BU represent the two directions of the rumor tree: Propagation (Top-Down) and Dispersion (Bottom-Up), as shown in Fig. 1A), we adopt the same encoder architecture for the bi-directional structure. First, we encode the ith sample as follows: (6)${\displaystyle L_{1,i}=LeakyReLU\left(W_1^T\left(A_i{X}_i\right)+b_1\right)}$ (7)${\displaystyle {\tilde{L}}_{1,i}=CONCAT\left(L_{1,i},X_i^{root}\right).}$

In this process, L_1,i represents the hidden layer features of the sample after the first layer of graph convolution, and $W_1^T$ and b₁ are parameters of the neural network. LeakyReLU is used here as the activation function. The feature vector of each node after encoding is subsequently concatenated with the feature vector R_i ∈ ℝ^d×1 of the root node of the sample before encoding, which is used to augment the features of the original post of the rumor event. $X_i^{root}={\left[R_i,R_i,\dots ,R_i\right]}^T$, $X_i^{root}\in {\mathbb{R}}^{n\times d}$ serves as the root node feature matrix to be concatenated for the ith sample, d is the node feature dimension. Afterward, ${\tilde{L}}_{1,i}$ is convolved multiple times to extract sample information, similar to the aforementioned process: (8)${\displaystyle L_{2,i}=LeakyReLU\left(W_2^T\left(A_i{\tilde{L}}_{1,i}\right)+b_2\right)}$ (9)${\displaystyle {\tilde{L}}_{2,i}=CONCAT\left(L_{2,i},L_{1,i}^{root}\right)}$ (10)${\displaystyle L_{3,i}=LeakyReLU\left(W_3^T\left(A_i{\tilde{L}}_{2,i}\right)+b_3\right)}$ (11)${\displaystyle L_{g_i}={\tilde{L}}_{3,i}=CONCAT\left(L_{3,i},L_{2,i}^{root}\right).}$

After passing the ith sample g_i through the same encoder L, we can obtain the bi-directional graph encoding representation of g_i as $L_{g_i}=\left(L_{g_i}^{TD},L_{g_i}^{BU}\right)$. A multi-level contrast is then performed, taking into account both the property augmentation of the local structure and the global topological high-level information.

For the node-level contrastive loss, we consider each direction of the bi-directional graph separately, following the steps below. For sample g_i, we construct positive instance pairs using the encoded original graph L_gi and its augmented graph L_aug(gi) as: (L_gi, L_aug(gi))⁺. Any distinct samples form negative instance pairs: (L_AGi, L_AGj)⁻. We use ι_node to denote the node-level contrastive loss. The contrastive loss for the positive instance pair of g_i is defined as ${\iota }_{node}^{pos}\left(g_i\right)$, and the contrastive loss for the negative instance pair is defined as ${\iota }_{node}^{neg}\left(g_i\right)$. Here, sim(⋅, ⋅) represents the similarity between different sets of sample encodings, which is measured by constructing the Cartesian product of the two sets and using cosine similarity to measure the distance between each pair. The sum of the resulting values represents the similarity between the two sets. (12)${\displaystyle {\iota }_{node}^{pos}\left(g_i\right)=-log\left(\frac{exp\left(sim\left(L_{g_i},L_{aug\left(g_i\right)}\right)/\tau \right)}{\sum _{j=1}^{n}exp\left(sim\left(L_{A{G}_i},L_{A{G}_j}\right)/\tau \right)}\right)}$ (13)${\displaystyle {\iota }_{node}^{neg}\left(g_i\right)=\frac{\sum _{i\ne j,j=1}^{n}exp\left(sim\left(L_{g_i},L_{A{G}_j}\right)/\tau \right)}{\sum _{j=1}^{n}exp\left(sim\left(L_{A{G}_i},L_{A{G}_j}\right)/\tau \right)}.}$

Inspired by infoNCE (Oord, Li & Vinyals, 2018), each contrast loss can be divided into upper and lower parts. The goal of ${\iota }_{node}^{pos}\left(g_i\right)$ is to increase the similarity between L_gi and L_aug(gi), while decreasing the similarity between different graphs L_AGi and L_AGj. The τ is a temperature parameter that is often used to modulate the sharpness of the probability distribution obtained from the similarity scores. Also, ${\iota }_{node}^{neg}\left(g_i\right)$ denotes a more exclusive encoding between g_i and AG_j in the context of decreasing the similarity between different graphs L_AGi and L_AGj.

By summing the contrastive losses of all samples, we obtain the final node-level contrastive loss ι_node. The smaller this loss, the closer the distance between the encodings of sample g_i before and after augmentation, and the farther the distance between the encodings of g_i and other distinct samples: (14)${\displaystyle {\iota }_{node}=\frac{1}{2N}\sum _{i=1}^n\left({\iota }_{node}^{pos}\left(g_i\right)+{\iota }_{node}^{neg}\left(g_i\right)\right).}$

For the graph-level contrastive loss, given the bi-directional graph encoding representation of g_i as $L_{g_i}=\left(L_{g_i}^{TD},L_{g_i}^{BU}\right)$, we use the mean pooling operator to aggregate information from both sets of encoding representations: (15)${\displaystyle S_{g_i}=CONCAT\left(MEAN\left(L_{g_i}^{TD}\right),MEAN\left(L_{g_i}^{BU}\right)\right).}$

Similar to the node-level contrastive loss, to pool the augmented samples, we obtain S_aug(gi). The goal is to minimize the distance between S_gi and S_aug(gi), while maximizing the distance between different samples and their augmented graphs. We define the contrastive loss for the positive instance pairs as ${\iota }_{graph}^{pos}\left(g_i\right)$, and for the negative instance pairs as ${\iota }_{graph}^{neg}\left(g_i\right)$: (16)${\displaystyle {\iota }_{graph}^{pos}\left(g_i\right)=-log\left(\frac{exp\left(sim\left(S_{g_i},S_{aug\left(g_i\right)}\right)/\tau \right)}{\sum _{j=1}^{n}exp\left(sim\left(S_{A{G}_i},S_{A{G}_j}\right)/\tau \right)}\right)}$ (17)${\displaystyle {\iota }_{graph}^{neg}\left(g_i\right)=\frac{\sum _{i\ne j,j=1}^{n}exp\left(sim\left(S_{g_i},S_{A{G}_j}\right)/\tau \right)}{\sum _{j=1}^{n}exp\left(sim\left(S_{A{G}_i},S_{A{G}_j}\right)/\tau \right)}}$

we can compute ι_graph: (18)${\displaystyle {\iota }_{graph}=\frac{1}{2N}\sum _{i=1}^n\left({\iota }_{graph}^{pos}\left(g_i\right)+{\iota }_{graph}^{neg}\left(g_i\right)\right).}$

Finally, by combining the node-level contrastive loss ι_node with the graph-level contrastive loss ι_graph, our multi-level contrastive loss can be represented as: (19)${\displaystyle \iota ={\iota }_{node}+\lambda {\iota }_{graph}}$

where λ is a parameter to be adjusted in order to balance the importance of ι_node and ι_graph.

Discriminator-based rumor detection

After multiple rounds of pre-training to minimize the multi-level contrastive loss, we obtain a high-performance encoder. This encoder can be used to encode a rumor event, which is represented as: (20)${\displaystyle L_{g_i}=CONCAT\left(L_{g_i}^{TD},L_{g_i}^{BU}\right).}$

We employ a multi-layer perceptron (MLP) for rumor detection classification. In addition to using the encoding L_gi obtained from the model, we also utilize the root feature of the rumor event $X_i^{root}$ and the graph-level representation MEAN(g_i) of the original rumor tree as additional information. Ultimately, we use these three parts as inputs and employ the MLP and softmax to perform classification. (21)${\displaystyle I_i=CONCAT\left(L_{g_i},X_i^{root},MEAN\left(g_i\right)\right)}$ (22)${\displaystyle \widehat{y_i}=softmax\left(MLP\left(I_i\right)\right).}$

Then, we use the cross-entropy loss function to measure the distance between the predictions $\hat{y}$ and the ground truth y for all rumor events: (23)${\displaystyle \mathcal{L}\left(y,\hat{y}\right)=-\sum _{i=1}^N{y}_{i}log\left({\hat{y}}_i\right)+\lambda \left|\right|\Theta |{|}_2^2.}$

In the loss function, $\left|\right|\Theta |{|}_2^2$ represents the L2 regularization term for the model parameters, and λ is a balancing factor.

Experimental settings

Experimental setup

We make comparisons with the following state-of-the-art baselines:

• Decision tree classification (DTC) (Castillo, Mendoza & Poblete, 2011): A decision tree model that combines various news features.

• Support Vector Machine based on Time Series (SVM-TS) (Ma et al., 2015): A linear SVM classifier that builds a time-series model using custom features.

• BU-RVNN (Ma, Gao & Wong, 2018): The structural recursive neural model based bottom-up tree neural network (with GRU unit) is used by the rumor detection method to learn about rumor information.

• TD-RVNN (Ma, Gao & Wong, 2018): The structural recursive neural model based top-down tree neural network (with GRU unit) is used by the rumor detection method to learn about rumor information.

• Graph-aware Co-Attention Networks (Lu & Li, 2020): The rumor detection method proposes a GCN-based model that can characterize the rumor propagation mode and use the dual co-attention mechanism to capture the interaction between source text, user attributes, and propagation path.

• BiGCN (Bian et al., 2020): The rumor detection approach uses a GCN-based model to represent the overall structure of the rumor tree by utilizing its two primary characteristics, rumor propagation and dispersion.

• RDEA (He et al., 2021): A GNN model based on contrastive learning, which uses data augmentation to enable the model to do better at learning rumor propagation patterns.

• UMLARD (Chen et al., 2022): A multi-view learning rumor detection model proposed with a distinguishable feature fusion mechanism, acquiring user-aspect features, and incorporate them with content features.

• TISN (Luo et al., 2022): A rumor detection method considers time information and propagation structure, which Transformer encoder is used to extract text information, and GCN is used to learn the patterns of rumor propagation.

To provide a fair comparison, we randomly divided the datasets into five sections and performed 5-fold cross-validation to achieve robust results. We assess the accuracy (Acc.) over the two categories and the F1 measure (F1) for each class for the two Twitter datasets.

In the experiment, BiMGCL uses the Adam optimizer (Kingma & Ba, 2014). The feature length of each node in the feature matrix, drop rate in edge modification, and dropout rate are set to 64, 0.4, and 0.5, respectively. The epoch of self-supervised training is set to 25, while the supervised fne-tuning process is iterated upon 100 epochs, and early stopping (Yao, Rosasco & Caponnetto, 2007) is applied when the validation loss stops decreasing by 10 epochs. The software versions we use are Python-3.7, PyTorch-1.8.1 (CPU), pytorch_geometric-1.7.0. In the graph construction phase we use the APIs provided by pytorch_geometric. In the data augmentation phase, both the topology and node attributes of the graph are augmented to produce richer samples. The code for our implementation can be found in https://github.com/vivian-166/BiMGCL/ and we provide a link to raw data: https://github.com/vivian-166/BiMGCL/tree/main/data.

Results and analysis

In this section, we evaluate the performance of the proposed framework for rumor detection by comparing it with the baseline methods. According to Tables 2 and 3, which display the experimental findings, our proposed BiMGCL beats all baseline approaches in terms of all assessment measures. In actuality, the following characteristics are what led to BiMGCL’s improvement.

First, our proposed BiMGCL not only models various rumor post objects and their comments as graph structures to obtain information about the contextual structure of their propagation patterns, but also considers propagation and dispersion as two crucial characteristics of rumors, which facilitates the learning of more fine-grained propagation patterns. This is in contrast to feature-based supervised learning rumor detection methods such as SVM-TS and DTC. In other words, these structural and semantic characteristics, which were discovered from the bi-directional graph of rumors, aid in enhancing the discriminability of rumor events, which results in the desired performance of BiMGCL. It is worth noting that our proposed BiMGCL improves 1.2%–42.8% over the feature-based supervised learning approach in terms of accuracy.

Second, to train the model in a supervised manner, most existing graph-based supervised learning rumor detection algorithms rely on a large corpus of labeled data (Sun et al., 2022; Wei et al., 2023). However, due to weak generalization for new target classes, such techniques fall short of detecting out-of-sample rumor events. In particular, GCAN, BiGCN, UMLARD and TISN employing graph neural networks to learn low-dimensional representations from rumor propagation trees, which typically pose a serious detection performance when the availability of training samples or labeling information is limited. Although UMLARD and TISN combined more information about rumors, they did not consider the relationship between different rumor events. Our proposed BiMGCL, in contrast to existing approaches, shifts the emphasis from labeling information to learning more sophisticated and discriminative patterns of rumor information propagation using multi-level contrastive learning. As a result, BiMGCL demonstrates improved rumor detection capabilities.

Third, BiMGCL has two advantages over the most relevant self-supervised work RDEA. On the one hand, RDEA investigates oversimplified data augmentation for node-to-node comparison learning, which applies to undirected graphs because they ignore nonlinear and hierarchical dependencies between rumor propagation patterns, leading to poor performance in out-sample rumor detection. In contrast, BiMGCL designs three insightful bi-directional graph data augmentation methods to identify more advanced rumor information, leveraging the rumor propagation and diffusion discriminant patterns, combining semantic and structural priors to increase the efficiency of rumor detection. Besides, compared to the node-to-node comparison learning approach in RDEA, we design a more expressive multi-level comparison approach to help graph encoders extract rumor knowledge from historical representations, which not only provide a stronger regularization to our bootstrapping objective but also enrich the self-supervision signals during the optimization.

Furthermore, the originating post of a rumor event often contains a wealth of detailed information and can wield substantial influence. However, UMLARD and TISN tend to overlook the crucial value embedded in the initial post’s information during the propagation process. Conversely, BiMGCL stands out for its adeptness at harnessing the data present in the source post. This enables BiMGCL to thoroughly excavate the structural nuances within rumor trees, thereby significantly enhancing the efficacy of rumor detection.

Ablation experiments

We further conducted ablation experiments to analyze the effect of each variant of BiMGCL, investigating the effect of each module in the pretraining encoder.

The empirical results are summarized in Fig. 2. -O, -R, -B, -G, and -N represent our model without concatenating the original graph, without concatenating the root feature, without using bi-directional graph, without graph-level contrastive loss, and without node-level contrastive loss, respectively. We make the following three observations. First, the performance of BiMGCL decreases if any of these four components are missing, indicating that they are all crucial for BiMGCL. Second, the performance drop is the largest for -R, suggesting that root features are indispensable, and the source post plays an essential role in rumor detection. At the same time, the -B performance decreases significantly, which implies the importance of considering both top-down representation and bottom-up representation.

In addition to the ablation experiments on the methods used in the comparative learning model, we further conduct experiments on the data augmentation methods to explore their effectiveness. The results of the experiments are shown in Fig. 3, where ‘NO AUG’ represents that no data augmentation method is used, ‘AM’, ‘SS’, and ‘EM’ represent attribute masking, subsampling, and edge modification, respectively, which are the three data augmentation methods described in the above article, and ‘ALL AUG’ indicates that the three methods are used at the same time. It can be observed that when no data augmentation methods are used, the lack of training samples and insufficient data features have a greater impact on model training and fail to achieve better accuracy. This problem is partially alleviated when using these methods alone, and better results can be achieved by using all three augmentation methods together. This is because the data augmentation methods, on the one hand, expand the existing data and produce a large amount of data from the original graph, which enables the comparative learning discriminator to better distinguish the features between similar and different rumor graphs; on the other hand, these three data augmentation methods augment the data in terms of the graph features, graph localization information, and graph structure, respectively, and produce rumor graph samples that contain various semantics, which provides important supports for model training.

Early rumor detection

A crucial method of assessing the model is early rumor detection, which aims to identify rumors before they become widespread and have negative social effects. To build the early detection task, we set a series of detection deadlines and only use posts that were published before the deadlines to assess the accuracy of the proposed and baseline methods and confirm whether the model can accurately identify rumors based on the scant information carried by the current early moments.

The early rumor detection task is illustrated in Fig. 4 using the results of our BiMGCL method and several benchmarks. The BiMGCL method described in this research achieves a high level of accuracy very quickly after the source post-original broadcast. BiMGCL performs significantly better than the other models at each deadline, indicating that our method is not only beneficial for long-term rumor detection, but also more beneficial for early detection of rumors. Notably, our BiMGCL model outperforms other benchmark tests and is outstanding and consistent at all times, demonstrating the effectiveness of the combination of a bi-directional graph model and multi-level contrast learning for robust and accurate identification.