DAEDCMD: Continual Multimodal Misinformation Detection

• DAEDCMD is a continual multimodal misinformation detection framework that tackles catastrophic forgetting and future distribution shifts.
• It integrates a base feature extractor, a Dirichlet process-driven mixture-of-experts, and a neural ODE-based dynamics model to adapt and predict evolving misinformation patterns.
• Empirical results on datasets like GossipCop, Weibo, and Twitter show notable gains in Accuracy and Macro-F1 compared to static and other continual learning methods.

DAEDCMD is a continual multimodal misinformation detection method introduced for online data streams in which event-indexed datasets arrive sequentially and a single detector must be updated without revisiting past data. It is designed to address two coupled failure modes in continual-MMD: past knowledge forgetting, in which performance on earlier events degrades after adaptation to later ones, and future distribution shifts, in which evolving social environments alter the feature distributions of upcoming events. The method combines a base multimodal feature extractor, a dynamically adapted mixture-of-experts structure governed by a Dirichlet process, and a continuous-time environmental dynamics model based on neural ODEs to preserve prior capabilities while anticipating future event distributions (Wang et al., 8 Jul 2025).

1. Problem setting and scope

DAEDCMD operates in the continual-MMD setting, where a stream of event-indexed multimodal datasets $D_1,\ldots,D_K$ is observed sequentially and each $D_k$ contains text, image, and label pairs for event $k$ (Wang et al., 8 Jul 2025). The detector $F_\theta(\cdot)$ is updated on $D_1$, then $D_2$, and so forth, without revisiting past data. This formulation differs from conventional offline multimodal misinformation detection, whose detectors are supervised on collected static datasets and are therefore vulnerable to becoming outdated when new events emerge.

Within this setting, the paper identifies two central challenges. The first is catastrophic forgetting, described as past knowledge forgetting: when the model is fine-tuned on $D_k$, its performance on previous $D_j$ for $j<k$ degrades. The second is future distribution shift: the social environment evolves over time, so the feature distributions associated with future events drift as topics, styles, and misinformation tactics change. DAEDCMD is explicitly constructed to address both problems jointly rather than treating them as independent subproblems (Wang et al., 8 Jul 2025).

A common misconception is to treat continual-MMD as ordinary domain adaptation or as standard continual learning with unimodal inputs. The formulation used here is stricter: it requires a unified multimodal detector to maintain competence over prior events while adapting to new events and while modeling temporal evolution in the environment. This suggests that DAEDCMD should be understood as a task-specific continual learning system for multimodal veracity prediction rather than as a generic replay or regularization method.

2. Core architecture

DAEDCMD comprises three modules: a base feature extractor, a dynamically adapted mixture-of-experts module, and a continuous-time environmental dynamics model (Wang et al., 8 Jul 2025). The architecture is organized so that multimodal alignment, event-specific specialization, and temporal forecasting of feature distributions are handled by separate but interacting components.

The base feature extractor uses a text encoder $\theta^t$, exemplified by BERT, to produce $D_k$0, and an image encoder $D_k$1, exemplified by ResNet, to produce $D_k$2. A contrastive pre-alignment objective is applied to align the text and image representations: $D_k$3 The aligned representations are then fused as

$D_k$4

with fusion implemented by self-attention or a simple linear projection (Wang et al., 8 Jul 2025).

The second module is a dynamically adapted mixture-of-experts. Its purpose is to isolate event-specific parameters so as to reduce interference across events while retaining shared knowledge. The expert pool contains one event-shared expert $D_k$5 and a growing set of event-specific experts $D_k$6. Each expert $D_k$7 includes a discriminator part $D_k$8, implemented as a LoRA-style low-rank adapter, and a variational generator part $D_k$9, implemented as a VAE with encoder and decoder (Wang et al., 8 Jul 2025).

The third module is a continuous-time environmental dynamics model. It learns how the distribution of fake-news embeddings evolves over event time $k$0, and then samples features from predicted future distributions. This component distinguishes DAEDCMD from continual methods that only attempt to retain past knowledge. In the DAEDCMD formulation, anticipation of future environments is not auxiliary; it is part of the main predictive pipeline.

3. Dirichlet-process expert expansion and routing

Expert expansion in DAEDCMD is governed by a Dirichlet Process Mixture defined over batched features $k$1. For existing experts $k$2,

$k$3

and for a potential new expert $k$4,

$k$5

where $k$6 controls sensitivity to novelty (Wang et al., 8 Jul 2025).

In implementation, the model works with negative logs: $k$7

$k$8

with

$k$9

If $F_\theta(\cdot)$0, a new expert $F_\theta(\cdot)$1 is spawned and initialized from $F_\theta(\cdot)$2 (Wang et al., 8 Jul 2025).

Feature extraction is then performed through a scalar routing gate: $F_\theta(\cdot)$3

$F_\theta(\cdot)$4

The shared expert is updated by exponential moving average,

$F_\theta(\cdot)$5

This design is intended to preserve stable shared structure while permitting event-specific specialization (Wang et al., 8 Jul 2025).

A plausible implication is that the Dirichlet process serves two roles simultaneously: it allocates capacity when novel events are encountered and provides a mechanism for parameter isolation. The paper’s ablation results support this interpretation, since disabling Dirichlet expansion degrades performance.

4. Continuous-time environmental dynamics

For fake-article embeddings $F_\theta(\cdot)$6, DAEDCMD computes empirical Gaussian parameters at each event $F_\theta(\cdot)$7: $F_\theta(\cdot)$8

$F_\theta(\cdot)$9

These statistics are then evolved through neural ODEs,

$D_1$0

after encoding the initial states as

$D_1$1

Integration is performed with dopri5: $D_1$2

$D_1$3

The decoded predictions are

$D_1$4

The dynamics loss is

$D_1$5

At inference for event $D_1$6, DAEDCMD samples

$D_1$7

and appends the sampled feature to $D_1$8 (Wang et al., 8 Jul 2025).

This mechanism is specific to the distribution of fake-news embeddings rather than to the full data distribution. That modeling choice implies that DAEDCMD emphasizes the temporal evolution of misinformation patterns themselves. The paper describes this as learning how the distribution of fake-news embeddings drifts over event time, then sampling features from predicted future distributions (Wang et al., 8 Jul 2025).

5. Learning objective and online optimization

DAEDCMD optimizes four losses: contrastive alignment $D_1$9, variational generator regularization $D_2$0, dynamics loss $D_2$1, and veracity prediction $D_2$2. The veracity prediction term is

$D_2$3

The full objective is

$D_2$4

with hyperparameters typically set to $D_2$5, $D_2$6, and $D_2$7 (Wang et al., 8 Jul 2025).

The online update procedure processes the event stream $D_2$8 sequentially. The model initializes the encoders $D_2$9, the shared expert $D_k$0, an empty event-specific expert set, and the dynamics parameters $D_k$1. For each event $D_k$2, training proceeds for $D_k$3 iterations with early stopping on macro-F1. In each batch, the method extracts $D_k$4 and $D_k$5, computes $D_k$6, evaluates responsibilities $D_k$7, spawns a new expert if the novelty condition is met, computes $D_k$8, routes features via $D_k$9, updates the shared expert $D_j$0 by EMA, computes empirical $D_j$1, solves the ODE forward to obtain $D_j$2, samples $D_j$3, computes $D_j$4, and backpropagates the total loss (Wang et al., 8 Jul 2025).

The reported implementation choices are specific: BERT layers use learning rate $D_j$5, while the MoE, dynamics model, and classifier use $D_j$6; batch size is 32; early stopping uses macro-F1 with patience 5; the Dirichlet new-expert sensitivity $D_j$7 is tuned per dataset and is approximately 1; and the ODE solver is dopri5 from torchdiffeq (Wang et al., 8 Jul 2025). Ablation analysis indicates that disabling Dirichlet expansion, removing the shared expert, or dropping $D_j$8 all degrade performance.

6. Empirical evaluation and benchmarks

The evaluation uses three datasets: GossipCop with 12.8k pairs, Weibo with 9.5k, and Twitter with 13.9k, each split into four event subsets by single-pass clustering (Wang et al., 8 Jul 2025). The baselines include six static MMD methods—Base, SAFE, MCAN, CAFE, BMR, and GAMED—and three continual schemes: +EWC, +Replay, and +LoRAMoE.

Using GAMED as the backbone, DAEDCMD yields the following reported improvements:

Dataset	Accuracy	Macro-F1
GossipCop	$D_j$9	$j<k$0
Weibo	$j<k$1	$j<k$2
Twitter	$j<k$3	$j<k$4

Against the best continual learner, LoRAMoE, DAEDCMD adds approximately $j<k$5 to $j<k$6 points in Accuracy and approximately $j<k$7 to $j<k$8 points in F1 (Wang et al., 8 Jul 2025).

The ablation study on GAMED reports average drops when removing the environment feature $j<k$9 ($\theta^t$0 Acc), Dirichlet process expansion ($\theta^t$1 Acc), and the shared expert $\theta^t$2 via EMA ($\theta^t$3 Acc) (Wang et al., 8 Jul 2025). The forgetting curves verify that DAEDCMD’s $\theta^t$4 routing and expert isolation slow performance decay on $\theta^t$5, while the dynamics model case studies show better generalization on future events.

These results support a specific empirical interpretation: the gains are not attributed solely to larger model capacity or to a generic continual learning scaffold, but to the conjunction of expert isolation and environmental prediction. The paper’s own breakdown associates performance degradation with the removal of either component.

7. Significance, interpretation, and limitations of scope

DAEDCMD is significant within continual-MMD because it frames misinformation detection as a sequential, temporally nonstationary multimodal learning problem rather than as a static supervised classification task (Wang et al., 8 Jul 2025). Its defining contribution is the explicit coupling of memory-preserving specialization through a Dirichlet process-based mixture-of-experts with future-oriented distribution modeling through a continuous-time neural ODE.

The method also clarifies an important distinction between preserving old knowledge and preparing for new environments. Many continual learning approaches focus on the first objective alone through replay, regularization, or adapter isolation. DAEDCMD adds a second axis by estimating how fake-news embedding distributions evolve over event time. This suggests a broader methodological view in which continual misinformation detection benefits from both retrospective parameter protection and prospective feature modeling.

Its reported evidence remains bounded by the experimental scope described in the source. The datasets are GossipCop, Weibo, and Twitter, each divided into four event subsets by single-pass clustering, and the comparisons are against six MMD baselines and three continual schemes (Wang et al., 8 Jul 2025). Therefore, claims about generality beyond those settings would require additional evidence. A plausible implication is that DAEDCMD is best regarded as a specialized continual-MMD framework whose architectural principles may transfer to other nonstationary multimodal classification problems, but such transfer is not directly established by the reported experiments.

In summary, DAEDCMD is a method for continual multimodal misinformation detection that addresses catastrophic forgetting through a Dirichlet-expanded mixture-of-experts and addresses future distribution shift through a continuous-time environmental dynamics model. Within the reported benchmarks, this combination yields consistent gains in Accuracy and Macro-F1 over both static MMD systems and prior continual learning baselines (Wang et al., 8 Jul 2025).