E-CaTCH: Event-Centric Misinformation Detection

• E-CaTCH is a supervised, event-centric framework that detects multimodal misinformation on social media by integrating text, images, timestamps, and event dynamics.
• It constructs pseudo-events using textual similarity and temporal proximity, enabling adaptive cross-modal attention and soft gating for robust feature fusion.
• The framework employs trend-aware LSTM temporal modeling with adaptive class weighting, achieving high performance on datasets like Fakeddit and COVID-19 MISINFOGRAPH.

Searching arXiv for the exact E-CaTCH paper and closely related misinformation-detection context. E-CaTCH, short for Event-Centric Cross-Modal Attention with Temporal Consistency and Class-Imbalance Handling, is a supervised framework for multimodal misinformation detection on social media. It models each post as a tuple containing text, image, timestamp, and binary label, but departs from post-independent classification by treating misinformation as an event-level, temporally evolving phenomenon. When event annotations are unavailable, it constructs pseudo-events from textual similarity and temporal proximity; within each event, it encodes text and images, aligns them through bidirectional cross-modal attention, fuses them with a soft gating mechanism, and models narrative progression with overlapping time windows and a trend-aware LSTM. Training combines adaptive class weighting, temporal consistency regularization, and hard-example mining, with reported gains on Fakeddit, IND, and COVID-19 MISINFOGRAPH, as well as cross-dataset robustness (Mousavi et al., 15 Aug 2025).

1. Problem formulation and event-centric perspective

E-CaTCH addresses binary misinformation detection for posts of the form

$(\mathbf{s}_i, \mathbf{i}_i, y_i, t_i),$

where $\mathbf{s}_i$ is text, $\mathbf{i}_i$ is image content, $y_i \in \{0,1\}$ is the label, and $t_i$ is the timestamp. The framework is motivated by four coupled difficulties: multimodal inconsistency, temporal drift, latent event structure, and class imbalance. In this formulation, a misleading post may arise not from either modality in isolation, but from their interaction; similarly, a narrative may become deceptive only through its temporal evolution rather than through any single post (Mousavi et al., 15 Aug 2025).

The framework’s defining move is to recast misinformation detection from instance-level classification to event-centric modeling. Posts that belong to the same rumor cycle, topic, or discourse phase are processed jointly, because the model assumes that misinformation dynamics are better aligned with evolving events than with independent observations. This shifts the unit of reasoning from isolated multimodal items to grouped trajectories of posts over time.

A central implication is that event structure becomes part of the predictive signal. This suggests that E-CaTCH is not merely a multimodal encoder with temporal smoothing, but a structured detector in which representation learning, fusion, and classification are conditioned by event membership. The paper therefore positions event construction as a foundational preprocessing and modeling step rather than as an optional augmentation.

2. Pseudo-event construction and temporal segmentation

When event labels are unavailable, E-CaTCH clusters posts into pseudo-events using textual similarity and temporal proximity. For Fakeddit, the paper gives the clearest procedure: posts are encoded with BERT-based sentence embeddings and grouped by agglomerative hierarchical clustering using cosine similarity as the distance metric. The number of clusters $|\mathcal{E}|$ is treated as a tunable hyperparameter selected empirically from clustering coherence and downstream classification performance (Mousavi et al., 15 Aug 2025).

The framework is less uniform across datasets than the name alone might suggest. For IND, event structure is approximated through consecutive daily windows with 50% overlap; for COVID-19 MISINFOGRAPH, it is approximated through weekly temporal windows with 50% overlap. For Fakeddit, by contrast, pseudo-events are explicitly constructed through text-driven clustering. This indicates that “event” is operationalized differently depending on annotation availability and corpus structure.

The paper is explicit that temporal proximity matters conceptually, but it does not provide a full joint clustering objective or a precise temporal-distance formula. It also does not specify linkage criteria, threshold schedules, or cluster-size constraints. A common misconception would be to read E-CaTCH as a fully formalized event-discovery model; the paper instead presents a pragmatic event-construction layer whose exact mechanics are only partially specified.

After events or event proxies are formed, posts are sorted by time and divided into overlapping windows. For Fakeddit, the windows typically span 3–5 days with 50% overlap. These windows are the units of temporal aggregation that feed the downstream sequence model.

3. Cross-modal representation learning and gated fusion

Within each post, text is encoded by BERT-base and image by ResNet-152, both projected into a shared $d$-dimensional space. The textual representation is derived from the BERT [CLS] embedding: $$\mathcal{T}_i = \mathbf{W}_{\text{text}}\, \text{BERT}_{\text{CLS}}(\mathbf{s}_i) + \mathbf{b}_{\text{text}},$$ with

$$\mathbf{W}_{\text{text}} \in \mathbb{R}^{d \times 768}, \qquad \mathbf{b}_{\text{text}} \in \mathbb{R}^{d}.$$

The visual representation is

$$\mathcal{I}_i = \mathbf{W}_{\text{img}}\, \mathbf{z}_i + \mathbf{b}_{\text{img}},$$

with

$\mathbf{s}_i$0

Missing images are represented as zero vectors to preserve tensor compatibility (Mousavi et al., 15 Aug 2025).

Each modality is refined with multi-head self-attention built from the standard scaled dot-product form

$\mathbf{s}_i$1

and the usual multi-head construction

$\mathbf{s}_i$2

The modality-specific refinements are written as

$\mathbf{s}_i$3

$\mathbf{s}_i$4

The paper states that standard Add & Normalize components, including LayerNorm, are intentionally omitted. It does not clearly specify whether attention operates over token/region sequences or over pooled post-level embeddings, so the exact granularity of intra-modal attention remains ambiguous.

Cross-modal alignment is then performed bidirectionally. Text attending to image is defined as

$\mathbf{s}_i$5

while image attending to text is

$\mathbf{s}_i$6

This directional asymmetry is central to the design: one modality may be reliable while the other is noisy, irrelevant, or deliberately misleading.

The two directional cross-modal representations are fused by a soft gating mechanism: $\mathbf{s}_i$7 where the gate is

$\mathbf{s}_i$8

with

$\mathbf{s}_i$9

This yields a contextualized post representation $\mathbf{i}_i$0. The paper presents the gate as an adaptive reliability mechanism: it learns whether text-conditioned image evidence or image-conditioned text evidence should dominate for each post.

4. Temporal modeling, semantic shift, and event representation

Inside each event, E-CaTCH models narrative evolution through overlapping temporal windows. If $\mathbf{i}_i$1 denotes a window, each post in that window receives a recency weight

$\mathbf{i}_i$2

where $\mathbf{i}_i$3 is a temporal decay coefficient. The window representation is then

$\mathbf{i}_i$4

This gives more influence to more recent posts within the window (Mousavi et al., 15 Aug 2025).

Temporal change is represented with two signals. The first is semantic shift,

$\mathbf{i}_i$5

which measures directional change between consecutive windows. The second is a momentum term,

$\mathbf{i}_i$6

with $\mathbf{i}_i$7, which tracks the persistence or acceleration of narrative movement. These are concatenated with the current window content: $\mathbf{i}_i$8

The temporal encoder is described as a trend-aware LSTM, but the trend-awareness comes from the augmented input rather than from a modified recurrent cell. The recurrence is

$\mathbf{i}_i$9

The final hidden state $y_i \in \{0,1\}$0 is used as the event representation. Classification is then performed by a sigmoid head: $y_i \in \{0,1\}$1

One ambiguity in the paper concerns the unit of supervision. The narrative description repeatedly states that classification is event-level, but the written classification loss uses per-instance notation $y_i \in \{0,1\}$2 and $y_i \in \{0,1\}$3. The most consistent reading is that E-CaTCH builds post-level multimodal embeddings but ultimately predicts at the event level through the final temporal state. The inconsistency is acknowledged in the paper’s technical details and is a genuine notational issue rather than a matter of interpretation.

5. Objective function, class imbalance handling, and reported results

E-CaTCH uses adaptive class weighting to counter label skew: $y_i \in \{0,1\}$4 Its event-wise weighted classification loss is written as

$y_i \in \{0,1\}$5

and aggregated over events as

$y_i \in \{0,1\}$6

Temporal smoothness is encouraged by a temporal consistency penalty

$y_i \in \{0,1\}$7

where

$y_i \in \{0,1\}$8

The total loss is

$y_i \in \{0,1\}$9

The model also uses hard-example mining, described as selecting the top-$t_i$0 highest-loss samples at each epoch, although the paper does not fully formalize the selection rule or its interaction with event grouping (Mousavi et al., 15 Aug 2025).

The reported results are as follows.

Dataset	Accuracy	F1
Fakeddit	$t_i$1	$t_i$2
IND	$t_i$3	$t_i$4
COVID-19 MISINFOGRAPH	$t_i$5	$t_i$6

For the same three datasets, the reported AUC-ROC values are $t_i$7, $t_i$8, and $t_i$9, respectively. On Fakeddit, the framework reports accuracy $|\mathcal{E}|$0, precision $|\mathcal{E}|$1, recall $|\mathcal{E}|$2, and F$|\mathcal{E}|$3 $|\mathcal{E}|$4. The paper states that E-CaTCH consistently outperforms state-of-the-art baselines on Fakeddit, IND, and COVID-19 MISINFOGRAPH, and that cross-dataset evaluations demonstrate robustness and generalizability. Reported transfer results include Fakeddit $|\mathcal{E}|$5 IND at $|\mathcal{E}|$6, Fakeddit $|\mathcal{E}|$7 COVID-19 at $|\mathcal{E}|$8, IND $|\mathcal{E}|$9 Fakeddit at $d$0, and COVID-19 $d$1 Fakeddit at $d$2.

6. Interpretability claims, implementation profile, and limitations

The framework is presented as both interpretable and scalable. Its interpretability claim rests on architectural transparency: pseudo-events expose event units, bidirectional cross-modal attention exposes directional dependencies, soft gates expose modality weighting, and temporal hidden states expose narrative evolution over overlapping windows. The paper does not, however, provide a dedicated quantitative interpretability study; the claim is therefore primarily structural rather than experimentally established (Mousavi et al., 15 Aug 2025).

On the systems side, E-CaTCH reports substantial efficiency. Using dual NVIDIA H100 GPUs, mixed precision FP16, efficient attention kernels, gradient checkpointing, and parallelization, it reports 98% GPU utilization, 989 TFLOPS, and a Fakeddit training time of approximately 14.8 hours. The official implementation is reported as publicly available at the paper’s repository.

Several limitations are explicitly or implicitly acknowledged. First, important implementation details are under-specified: the paper does not clearly provide the hidden dimension $d$3, number of attention heads $d$4, optimizer, learning rate, dropout, batch size, or exact values of $d$5, $d$6, $d$7, and $d$8. Second, the event-construction stage is central but only partially formalized, especially outside the Fakeddit setting. Third, the paper contains notation inconsistencies, most notably around event-level versus post-level classification. Fourth, the granularity of cross-modal attention is unclear: the equations suggest pooled feature interaction rather than explicit token-region alignment, but the paper does not settle the issue.

The paper also suggests likely failure cases associated with the design: poor pseudo-event clustering, weak or noisy timestamps, sparse events with too few windows, jointly misleading but internally consistent multimodal posts, and domain shifts in which text similarity no longer groups narratives reliably. Future extensions proposed in the paper include adding audio and video modalities, incorporating external knowledge bases and fact-checking resources, improving domain adaptation and continual learning, integrating social-contextual signals, and strengthening explainability through attention-over-time dashboards.

In the literature of multimodal misinformation detection, E-CaTCH is best understood as an event-structured, temporally aware fusion architecture. Its distinguishing contribution is not simply the use of attention, but the coupling of pseudo-event organization, bidirectional cross-modal alignment, adaptive gating, and trend-aware temporal encoding into a single supervised framework for misinformation dynamics (Mousavi et al., 15 Aug 2025).