Tracing truth: dynamic temporal networks for multi-modal fake news detection

Text feature representation

To obtain meaningful vector representations for diverse textual modalities, we employ a hybrid encoding strategy leveraging two pre-trained language models. Given an input sentence S, we first tokenize it with padding or truncation into a fixed-length sequence $T=\left\{\left[CLS\right],s_1,s_2,\dots ,s_i\dots ,s_n,\left[sep\right]\right\}$, where s_i denotes the token corresponding to the $i$-th word. This sequence is passed into a fine-tuned robustly optimized bidirectional encoder representations from Transformers (RoBERTa) encoder to extract contextualized embeddings. The hidden state of the special token $\left[CLS\right]$ is used as the sentence-level embedding, denoted by e ∈ ℝ^d_t, where d_t is the embedding dimension. In parallel, we employ a fine-tuned T5 encoder to process structured textual fields in the news data, including the title, content, and associated post. Each field is encoded independently to produce dense vector representations e^t, e^c, and e^p, respectively. These embeddings are further normalized to ensure consistency across modalities and facilitate downstream training. The RoBERTa-based representation is specifically used for user-related textual inputs, such as user descriptions or profile metadata, yielding the user embedding e^u. This dual-model setup enables the system to capture both general linguistic patterns and field-specific semantics across heterogeneous textual inputs.

Image feature representation

To incorporate visual modality into the model, we utilize a pre-trained ResNet18 network to extract semantic-level image features from news-related visual content. Given an input image associated with a news event, we pass it through ResNet18 and extract the feature vector from its final average pooling layer, resulting in a dense representation. This visual feature vector is then subjected to normalization to ensure alignment with the dimensions of other modalities. The resulting image embedding is denoted as e^v ∈ ℝ^d_v, where d_v indicates the dimension of the image feature vector. By integrating visual cues from news images, the model is able to capture multimodal signals that may reflect emotional tone, contextual clues, or visual bias, thereby enhancing its capacity for comprehensive news understanding and veracity assessment.

Social media data feature representation

When a news topic emerges, it inevitably triggers public opinion through social media, and metrics such as the number of reposts and likes on posts related to the topic are key numerical attributes that need to be considered. In particular, we need to focus on the social media data of the user u_i who posted the news and related posts. This data contains key attributes, including follower count, fan count, and the verification status of the user. These user features should be treated as important numerical attributes and further incorporated into the model for analysis. Subsequently, the aforementioned social media numerical features are converted into a sparse feature matrix using one-hot encoding, with the embedding representation of the user’s social media metrics denoted as e^u. In this way, the numerical attributes can be normalized. Quantifying these social media data not only helps to fully understand the user’s influence and credibility on the platform but also provides more accurate and comprehensive information support for the model, improving its prediction capability and effectiveness.

Definition of graph structure

As shown in Fig. 4, we model the relationships between news using a graph structure to further analyze and detect fake news propagation patterns. Drawing from the entity categories identified in the dataset, including users, news, and posts—we construct a heterogeneous news graph 𝔾 = {V, ϵ}, where V = {U, N, P} represents the set of nodes for users, news, and posts, respectively. $U=\left\{u_1,u_2,\dots ,u_i\dots ,u_{c_u}\right\},N=\left\{n_1,n_2,\dots ,n_i\dots ,n_{c_n}\right\},P=\left\{p_1,p_2,\dots ,p_i\dots ,p_{c_p}\right\}$ represent the sets of users, news, and posts, with c_u, c_n, c_p denoting the number of users, news, and posts, respectively. ϵ represents the edges that link these nodes. Specifically, we model news, users, and posts as nodes, with different types of interactions as edges, constructing a multi-type edge graph that reflects the propagation patterns in social networks. The implementation steps are outlined below:

Nodes: The graph nodes correspond to three categories of entities:

1. User nodes: Each user posting news is represented as a node in the graph.

2. News nodes: Each news is represented as a node.

3. Post nodes: Each post related to the news is represented as a node.

Edges: Based on different types of interactions between news, we define the following edges between nodes:

1. news-post interaction edges: Recursively add n − p edges between main news nodes and their directly replied posts. Add p − p edges between reply posts to represent interactions among posts.

2. user interaction edges: Add n − u or p − u edges between each main news or post and its corresponding user. If a post has replies, add u − u edges to represent interactions between users.

Heterogeneous graph optimization

After constructing a network graph based on different node types (users, news, and posts), as shown in Fig. 3, a random walk process is simulated between nodes to capture the underlying patterns of information propagation, going beyond the relationships of first-order neighbors to obtain a more comprehensive propagation graph. The specific steps are as follows:

1. Random walk initiation: Start from an initial node and iteratively visit its neighboring nodes according to the connections defined in the adjacency list.

2. Frequency sampling to select high-frequency neighbors: In large-scale networks, connections between nodes are often dense. To reduce complexity, a frequency sampling method is used to retain only high-frequency neighbors, controlled by parameter k. Neighbor nodes are ranked by visitation frequency during random walks, excluding the starting node and irrelevant nodes. High-frequency nodes are prioritized and retained in the final neighbor list.

3. Enforced retention of important edges: To ensure that certain critical edge types (e.g., news-user, post-user) are always retained, these edges are forcibly added to the neighbor list even if they are not part of the initial target set of the random walk. This guarantees that important connections are reflected in the graph structure, enhancing overall connectivity.

Temporal sequence penetration fusion module

As the internet and social media continue to evolve, focusing on the time dimension is crucial for studying news dissemination. By analyzing the temporal proximity of nodes, the concept of time similarity intensity is introduced to measure the impact of related news and posts on information spread. The core idea is that the closer the time, the stronger the correlation in information dissemination, enabling a more accurate assessment of the cumulative effects and diffusion paths of information sources.

i. Incorporating time information into attribute features and graph structure

As shown in Fig. 5, convert the time string (e.g., Mon Apr1218:37:46 2021) into a UTC timestamp to embed temporal information into the identifiers of graph nodes and edges, integrating time data into the graph structure. Node identifiers embed time by concatenating the news ID with the timestamp (e.g., news_id+^′′t^′′ + changetime(source_tweet[‘created_at′])). Similarly, edge identifiers incorporate reply hierarchy relationships and timestamps to associate news with posts or posts with other posts in the graph structure. Even if replies share the same news ID, they can be distinguished by their timestamps. In the final graph file, all relationships between nodes and edges include temporal information, ensuring that the graph structure’s relationships are closely linked to time.

ii. Enhancing the target news node with temporal similarity based on neighbor features

By leveraging neighbor features and time similarity, the target news node is enhanced to capture the dynamics of news dissemination in complex networks. Temporal similarity is determined by measuring the interval separating the x − th neighbor and the target news n_i in its neighborhood N^n_i and the target news n_i itself.

The time interval $I_x^n$ is calculated according to the following formula: ${\displaystyle I_x^n=\frac{exp\left(t_x-t_{n_i}\right)}{\sum _{j=1}^{k_n}exp\left(t_j-t_{n_i}\right)}.}$

Here, t_ni the publishing timestamp of the target news, t_x indicates the publication moment of the x − th neighbor within the neighborhood N^n_i associated with n_i, while k_n refers to the count of news neighbors associated with n_i. Nodes that are temporally closer exhibit stronger correlations in information dissemination and greater content relevance. Therefore, we assign higher weights to such nodes. The temporal similarity strength $S_n^x$ between a neighboring node in Nⁿ and the target news n_i is defined by the following formula: ${\displaystyle S_x^n=\frac{exp\left(-I_x^n\right)}{\sum _{j=1}^{k_n}exp\left(-I_j^n\right)}.}$

Here, $I_x^n$ represents the time gap between the publication of the x − th neighbor in Nⁿ and the target node n_i. k_n denotes the total count of neighbors associated with n_i.

Similarly, the temporal similarity strength S^p between posts can be calculated using the same method described above, where Sⁿ ∈ ℝ^k_n and S^p ∈ ℝ^k_p.

iii. Concatenating the attribute characteristics of news nodes

A self-attention mechanism uncovers feature dependencies, enhancing contextual semantics, First, the attribute features of the news node, including its title and content, are processed and combined. Then, they are linearly transformed to a unified dimension d, facilitating subsequent neural network computations. The title feature of the neighboring news E^t is given by the following formula: ${\displaystyle E^t=concat\left(e_1^t,e_2^t,\dots \dots ,e_{k_n}^t\right)W^t.}$

Here, W^t ∈ ℝ^d_t×d represents the trainable parameters, where d_t and d denote the unified projection dimension of the title feature and the embedding dimension, respectively. k_n the count of neighboring nodes associated with the target news n_i. concat(⋅) denotes the vector concatenation operation, and E^t ∈ ℝ^k_n×d.

Each embedding vector undergoes linear transformations through a self-attention layer to generate query, key, and value vectors. Three independent linear layers are initialized to perform matrix operations on the embedding vectors, producing tensors with a shape of [batch_size, seq_len, n_heads*head_dim] represents the vector dimension for each attention head. The output of the multi-head attention is further processed through a linear layer to produce tensors of the same shape. This operation integrates the attribute features from the news titles of neighboring nodes and employs a multi-head attention mechanism to refine the current news node’s attributes, enhancing its semantic dependencies and structural features. The semantic relational features of the news title $E_{rel}^t$ are calculated using the following formula: ${\displaystyle E_{rel}^t=MultiHead\left(Q,K,V\right)=MultiHead\left(E^t,E^t,E^t\right).}$

Here, $E_{rel}^t\in {\mathbb{R}}^{k_n\times d}$, the formula for $E_{rel}^t$ means that the input E^t is mapped into three matrices: query, key, and value, through the multi-head attention mechanism. Specifically, it is linearly transformed as $Q^{{}^{\prime }}=E^t{W}_t^Q$, $K^{{}^{\prime }}=E^t{W}_t^K$, and $V^{{}^{\prime }}=E^t{W}_t^V$, where $W_t^Q$, $W_t^K$, and $W_t^V$ represent the weight matrices corresponding to the query, key, and value. The multi-head attention mechanism is calculated using the following equation: ${\displaystyle MultiHead\left(Q,K,V\right)=\left(\underset{i=1}{\stackrel{h}{\parallel }}Attention\left(Q{W}^Q,K{W}^K,V{W}^V\right)\right)W^M}$ ${\displaystyle Attention\left(Q^{{}^{\prime }},K^{{}^{\prime }},V^{{}^{\prime }}\right)=softmax\left(\frac{Q^{{}^{\prime }}{K^{{}^{\prime }}}^T}{\sqrt{d_k}}\right)V^{{}^{\prime }}.}$

Here, d_k represents the dimension of K′. $W^Q\in {\mathbb{R}}^{d\times \frac{d}{h}}$, $W^K\in {\mathbb{R}}^{d\times \frac{d}{h}}$, and $W^V\in {\mathbb{R}}^{d\times \frac{d}{h}}$. d indicates the unified projection size, and —— signifies the operation of joining features. Using a comparable method described above, the semantic relational features of the news content $E_{rel}^c$, the visual semantic relational features $E_{rel}^v$, and the temporal semantic relational features $E_{rel}^{s^n}$ for neighboring nodes can also be obtained.

Time information from neighboring nodes is integrated into the attribute features to capture their dynamic relationships. Subsequently, varying weights are assigned to each neighbor. In real social networks, the process of news dissemination gives varying levels of attention to different neighboring nodes, enabling the dynamic capture of the social structure information of the news.

iv. Temporal weighted diffusion features

Finally, the fused feature representation for each node is obtained. The calculation formula for the temporal diffusion features of the news title is as follows: ${\displaystyle E_{tdf}^t=MultiHead\left(Q,K,V\right)=MultiHead\left(E_{rel}^{s^n},E_{rel}^t,E_{rel}^t\right).}$

Here, $E_{rel}^{s^n}$ and $E_{rel}^t$ denote the temporal semantic relationships and the title-related semantic relationships associated with the main news. $E_{tdf}^n\in {\mathbb{R}}^{k_n\times d}$ denotes the temporal diffusion attribute of the focus news. k_n denotes the count of neighboring nodes linked to the target news n_i. Using the above formula, we can obtain the temporal diffusion features of the news content $E_{tdf}^c$.

Based on the node types (news, user, post), we process the embeddings of different types of neighboring nodes, apply the attention mechanism, and fuse temporal and content information, followed by processing with a bidirectional recurrent neural network (Bi-RNN). The embeddings are then aggregated using mean pooling, and the pooled result is returned. Finally, the fused features of the news node are obtained, and the calculation formula is as follows: ${\displaystyle E_{fus}^n=Fusio{n}_{Bi-RNN}\left(E_{tdf}^t,E_{tdf}^c,E_{tdf}^c\right)=MeanPool\left(E_{tdf}^t,E_{tdf}^c,E_{tdf}^c\right).}$

Here, $E_{fus}^n\in {\mathbb{R}}^{k_n\times d}$, where k_n refers to the count of connected nodes related to the main news n_i. MeanPool(⋅) denotes the operation of mean pooling.

Integrating the graph structure enhancement module, we extract fused features for users, news, and posts in the heterogeneous graph. Using the temporal sequence permeation fusion module, we dynamically capture the contextual semantics of neighboring nodes within the social network, integrating time-related information. When calculating the fused features of post nodes, temporal permeation is not applied to user neighbors as there are no dynamic relationships between users. By combining semantically enhanced user features with other user attributes, the fused features of post nodes $E_{fus}^p$ are obtained. Finally, the fused features for target news nodes $E_{fus}^n\in {\mathbb{R}}^{1\times d}$, post nodes $E_{fus}^p\in {\mathbb{R}}^{k_p\times d}$, and user nodes $E_{fus}^n\in {\mathbb{R}}^{k_u\times d}$ are computed, where k_p and k_u indicate the quantities of post-type and user-type neighbors associated with n_i, and d refers to the dimensionality of the projection.

Dynamic monitoring in the temporal dimension

The above module optimizes fake news detection through complex networks and information dissemination. This subsection delves deeper from sociological and psychological perspectives. As shown in Fig. 6, fake news often draws inspiration from mainstream public opinion, using high-exposure content to attract attention, with more concentrated release and dissemination times. Monitoring the propagation disorder of rumors and non-rumors using entropy enhances the model’s detection accuracy and timeliness.

i. Calculation of dissemination disorder

Extract the source tweet’s publication time for each news item from the dataset as the starting point and record the reaction tweets’ publication times to form a time series. Define time intervals (e.g., 1 h, 2 h, 6 h, 12 h, or even up to 24 h, etc.) starting from the source tweet time, segment the timeline, and count the number of tweets in each interval to create a propagation distribution and calculate entropy. A smaller entropy indicates concentrated propagation, while a larger entropy suggests more dispersed or disordered propagation. ${\displaystyle H=-\sum _{i=1}^{k_p}p\left(x_{p_i}\right)logp\left(x_{p_i}\right).}$

Here, p(x_pi) represents the probability distribution of a specific post within each time period. k_p represents the count of post-type neighbors linked to n_i. By comparing the entropy changes of rumors and non-rumors, the differences in their propagation patterns can be analyzed. Low entropy indicates that rumors tend to spread rapidly in a short period, while high entropy reflects the sustained diffusion of real news across multiple nodes. In the early stages of propagation, the low number of tweets may cause entropy calculations to be affected by sparsity. This can be addressed using weighted entropy and smoothing techniques.

ii. Introducing disorder as a temporal feature into the model

Propagation entropy, as a measure of news dissemination disorder, can be combined with other features and input into the model to enhance news authenticity detection. Additionally, calculating entropy at different time intervals (e.g., every 6 or 12 h) forms a time series that can be input into a Transformer temporal model to capture propagation dynamics. This approach better reflects the timeliness and complexity of propagation paths, thereby optimizing the effectiveness of fake news detection.

Adding dynamic attention coefficients to optimize the graph structure

Temporal semantics are added to dynamic neighbor embeddings, and attention coefficients are computed considering node relationships. This means that the interaction between nodes depends not only on their static topology but also on their temporal interactions. The dynamic attention coefficient integrates temporal semantics, enabling the model to more accurately capture node dependencies and information flow at different time points. The formula for the dynamic attention coefficient δ₊ is as follows: ${\displaystyle {\delta }_{+}=LeakyReLU\left(\left(E_{fus}{W}_{{\alpha }_1}\right)+{\left(E_{fus}{W}_{{\alpha }_2}\right)}^T\right).}$

Here, E_fus represents the fused features of the news node, and W, α₁ and α₂ are trainable parameters. The adjacency matrix A is divided into upper d rows and lower d rows, corresponding to α₁ and α₂.αϵℝ^2d, E_fus ∈ ℝ^l×d, W ∈ ℝ^d×d, and δ₊ ∈ ℝ^l×l.

Subsequently, the dynamic attention coefficients are standardized using the following formula: ${\displaystyle {\delta }_{+}^{{}^{\prime }}=softmax\left({\delta }_{+}\right).}$

We use the softmax function for normalization because some posts express opposition to the news being released, which means that some values of δ₊ might be negative. However, after applying softmax brings values closer to zero, reducing the influence of negative ones. Let δ₋ =  − δ₊, and the normalization formula is as follows: ${\displaystyle {\delta }_{-}^{{}^{\prime }}=softmax\left({\delta }_{-}\right).}$

The graph attention mechanism performs a weighted calculation on the nodes in the embedding sequence by comprehensively analyzing the temporal context of each node and the dynamic relationships with its neighboring nodes. This dynamic weighting calculation not only reflects the immediate associations between nodes but also considers their interactions at different time steps, thereby improving the timeliness and accuracy of feature representation. By concatenating the aggregated features of both positive and negative correlations, the output features processed by the fully connected layer are represented as follows: ${\displaystyle f_g=ReLU\left(\left({\delta }_{+}^{{}^{\prime }}{E}_{fus}\left|\right|{\delta }_{-}^{{}^{\prime }}{E}_{fus}\right)W_g\right).}$

Here, E_fus represents the fused features of the news node, and W_g is a trainable parameter, with W_g ∈ ℝ^2d×d. ReLU(⋅) is used as a nonlinear function, || denotes a joining operation, and f_g ∈ ℝ^l×d.

During model training, different dynamic attention coefficients are produced. Based on the positive attention weights, the information from neighboring nodes is weighted and summed to generate updated representations for individual nodes. Similarly, the feature representation for negative attention is also derived. The node features under both positive and negative attention are concatenated to form a feature matrix that is twice the original size. A new weight matrix, used in a variant of the graph attention network (GAT), is then applied to the concatenated features for m linear transformations, generating new node features. This process yields a more comprehensive and integrated graph structure feature F_g. The calculation formula is as follows: ${\displaystyle F_g={\sigma }_{elu}\left(GAT\left(\underset{i=1}{\stackrel{m}{\parallel }}{f}_g^i\right)\right).}$

Here, $f_g^i$ represents the graph structural features obtained in the i − th iteration, where i ∈ [1, h].GAT(⋅) represents the graph attention module, ${\sigma }_{elu}\left(\cdot \right)$ stands for the elu activation function, || is the concatenation operation, and F_g ∈ ℝ^l×d.

Multimodal feature integration and prediction

This module aggregates the features of news nodes and their neighbors (users, posts) using a Transformer, integrating temporal features, node entropy weights, and fused visual-textual representations to enhance the perception of temporal dynamics and the complexity of multimodal node information. Utilizing the multi-head attention mechanism, it captures complex interactions between nodes and across modalities, generating enriched node representations for prediction. The classification process uses cross-entropy loss for training and leverages Adam for optimization.

1. Multimodal input and fusion strategy: The input includes multimodal data such as text, visual content, user information, and the level of news propagation disorder. Features from text and image modalities are fused through a visual-textual embedding mechanism, while user and propagation features are concatenated at the input stage. All modalities are then unified and weighted in the input layer to form a comprehensive feature representation for model training. The news propagation disorder feature is treated as an independent channel that captures abnormal dissemination patterns, and is integrated with other modality outputs at the decision layer to generate the final classification result.

2. Feature encoding and attention mechanism: The multimodal node features are further processed using a Transformer, incorporating node type and positional encoding. The multi-head attention mechanism dynamically assigns weights to different modality channels—including textual, visual, user profile, and propagation disorder features—automatically learning their relative importance for fake news detection. This facilitates the extraction of deep semantic dependencies across both content and structure, thereby enhancing model accuracy.

3. Prediction and classification: The final node embeddings are passed through a linear layer and an activation function, producing classification outputs for identifying fake news. The activation function is defined as:

The loss function uses binary cross-entropy combined with L2-normalized vectors to optimize the model and generate the final prediction probabilities for rumors or non-rumors.

Learning rate

The learning rate is a key hyperparameter that significantly impacts model training performance, including convergence speed, stability, and overall effectiveness. A larger learning rate can accelerate early parameter updates but may cause oscillations or miss the optimal solution, leading to non-convergence or degraded performance. A smaller learning rate ensures stable updates but slows down training and may get stuck in local optima, limiting performance improvement. In the experiments, the value of the learning rate α was assigned a value of 10⁻ⁱ, where i ∈ [1, 5]. Figure 7 presents the experimental results.

Overall, the results demonstrate that batch size significantly impacts DTN performance, and selecting an appropriate size based on dataset complexity and sample volume is essential to balance convergence speed and detection accuracy.

Batch size

In deep learning, batch size serves as an essential hyperparameter that impacts training efficiency. Varying batch sizes impact both gradient estimation accuracy and memory consumption. A smaller batch size enables more frequent parameter updates and introduces stochastic gradient estimates, which can help the model avoid local optima. However, excessively small batch sizes may result in instability, slowing down convergence and degrading final performance. Conversely, using larger batch sizes improves gradient estimation accuracy and training stability, but it reduces update frequency, resulting in longer durations and higher memory consumption.

In our experiments, the batch size was set as 2ⁱ, where i ∈ [2, 6]. The results, illustrated in Fig. 8, indicate that increasing the batch size improves DTN’s training speed significantly. However, when the batch size becomes too large, performance drops, likely due to reduced update frequency, causing the model to miss optimal convergence points. For the PHEME dataset, the highest accuracy of 0.938 is obtained with a batch size of 16, while the GossipCop dataset achieves its peak accuracy of 0.993 at the same batch size.

Overall, the results demonstrate that batch size significantly impacts DTN performance, and selecting an appropriate size based on dataset complexity and sample volume is essential to balance convergence speed and detection accuracy.

Overall model performance

Table 2 presents the comparative performance of a series of representative models on the PHEME and GossipCop datasets, with evaluation metrics covering accuracy, precision, recall, and F1-score for both real and fake news classification tasks. The best results for each metric are highlighted in bold. The visualized comparison results are shown in Figs. 9 and 10. The selected baselines encompass a diverse spectrum of methodological paradigms that collectively reflect the current landscape of fake news detection. Specifically, multimodal approaches such as SAMPLE, AMPLE, COOLANT, and MFAN exploit complementary textual and visual information to capture enriched semantic representations; graph-based models including TCGNN, MRHFR, and HCCIN emphasize relational and topological structures within content dissemination or user interaction networks; while hybrid architectures like HetTransformer and MIGCL integrate multimodal cues with graph-based reasoning to simultaneously leverage content semantics and structural dependencies. The inclusion of these diverse and competitive baselines not only facilitates a rigorous and multidimensional evaluation but also highlights the importance of jointly modeling multimodal semantics and structural context, which provides a comprehensive foundation for validating the effectiveness and generalizability of our proposed framework.

Performance on the PHEME dataset. On the PHEME dataset, the DTN model achieves the highest overall performance, with an accuracy of 0.938. For real news classification, it obtains precision, recall, and F1-score values of 0.967, 0.964, and 0.936, respectively, outperforming competitive baselines such as MFAN (0.997, 0.863, 0.925) and HCCIN (0.916, 0.930, 0.919).

Regarding fake news classification, DTN achieves precision, recall, and F1-score values of 0.895, 0.913, and 0.924, respectively, surpassing all other methods. Although MIGCL yields comparable results (0.908, 0.925, 0.917), DTN demonstrates a more balanced performance across metrics, reflecting its robustness in distinguishing both real and fake news.

Performance on the GossipCop dataset. On the GossipCop dataset, DTN achieves the highest accuracy of 0.993. For real news, it records precision, recall, and F1-score values of 0.992, 0.989, and 0.989, respectively. While HetTransformer attains a marginally higher precision of 0.994, its recall and F1-score (both at 0.993) remain comparable, and DTN exhibits greater consistency across both classes. For fake news classification, DTN again leads with precision, recall, and F1-score values of 0.982, 0.982, and 0.989, respectively. In comparison, MIGCL achieves scores of 0.964, 0.985, and 0.974, indicating that DTN maintains superior balance and overall effectiveness.

Summary. The results on both datasets confirm the superior performance of DTN compared to unimodal, multimodal, and hybrid baselines. While existing models such as MFAN, MIGCL, and HCCIN demonstrate strength in specific aspects, DTN consistently achieves competitive scores across all metrics. Its performance benefits from a unified framework that incorporates temporal dynamics, multimodal fusion, and structural reasoning, resulting in a more accurate, stable, and generalizable solution for fake news detection.