Dynamic Multi-Graph Fusion Module (DMF)

• DMF is a neural network module that dynamically fuses multiple heterogeneous graphs to produce refined node or graph-level embeddings.
• It employs techniques like attention-based fusion, dynamic gating, and time-decay weighting to adaptively aggregate diverse relational data, achieving notable performance gains.
• Applied in traffic forecasting, dynamic link prediction, and multimodal emotion recognition, DMF demonstrates improved robustness and computational scalability.

A Dynamic Multi-Graph Fusion Module (DMF) is a neural network component designed for adaptive, context-aware integration of multiple heterogeneous graphs—spanning modalities, feature perspectives, or discrete time snapshots—within graph-based learning pipelines. DMF modules enable models to dynamically combine information from several relational structures, addressing the limitations of single-view or static fusion schemes and delivering improved performance and robustness in domains such as traffic forecasting, dynamic link prediction, and multimodal emotion recognition (Rafi et al., 10 Jan 2026, Qi et al., 2024, Hu et al., 2022).

1. Foundational Principles

The primary objective of DMF is to generate node-level or graph-level embeddings that optimally aggregate the information present in multiple, potentially dynamic, graphs. Each graph encodes a distinct semantic, temporal, or modal view; for example:

• In spatiotemporal traffic applications, graphs may encode physical distance or dynamic travel times between nodes (Rafi et al., 10 Jan 2026).
• In temporal graph models, graphs may represent a sequence of historical snapshots, each corresponding to an interval in time (Qi et al., 2024).
• In multimodal learning, graphs may encode intra-modal (within-modality) and inter-modal (across-modality) relationships between features (Hu et al., 2022).

DMF modules employ mechanisms such as attention-based fusion, dynamic gating, or time-decay weighting to compute data-dependent aggregation of graph-derived embeddings, as opposed to static averaging or pre-specified rules.

2. Architectural Patterns and Module Instantiations

DMF designs exhibit considerable diversity depending on downstream tasks and data types:

• Spatiotemporal DMF (RL-DMF): Raw node features ($$\tilde X_{\mathrm{temp}}, \tilde X_{\mathrm{spatial}}$$) are processed via multiple GCNs, each parameterized by a dynamic adjacency ($\tilde{A}_t^g$) constructed from different physical or dynamical metrics (e.g., distance, travel time). The outputs are then adaptively fused using a learnable attention mechanism, producing embeddings $Z_t^{\mathrm{fused}}$ that serve as inputs to temporal models such as LSTMs (Rafi et al., 10 Jan 2026).
• Temporal Multi-Graph Fusion (SFDyG): DMF fuses a sliding window of discrete graph snapshots into a “temporal multi-graph” $\widetilde{\mathcal{G}}$, realizing time-aware edges via Hawkes process–driven decay factors. The fused weighted adjacency matrix is processed with GCN or GAT layers, yielding node embeddings that efficiently summarize history while mitigating computational overhead (Qi et al., 2024).
• Multimodal Dynamic Fusion (MM-DFN): Here, each utterance-modality pair defines a node in a multimodal graph. DMF comprises gated updates akin to LSTM cells, per-layer dynamic graph convolutions, and residual mixing, collectively regulating information exchange within and across modalities at each layer (Hu et al., 2022).

3. Mathematical Formulation

Spatiotemporal DMF (RL-DMF) (Rafi et al., 10 Jan 2026)

1. Input construction:

$$H_t = [\tilde X_{\mathrm{temp}}[:, t, :] \, \Vert \, \tilde X_{\mathrm{spatial}}]$$

2. Graph construction: For $K$ graphs (e.g., $K=2$; distance $G^d$ and travel-time $G^{tt}$), normalize each adjacency, add self-loops, and row-normalize:

$$\tilde{A}_t^g = D_t^{-\tfrac{1}{2}}(A_t^g + I)D_t^{-\tfrac{1}{2}}$$

3. Graph convolution:

$\tilde{A}_t^g$0

4. Fusion via attention:

$\tilde{A}_t^g$1

$\tilde{A}_t^g$2

Fused embeddings propagate to temporal LSTM layers.

Temporal Multi-Graph Fusion (SFDyG) (Qi et al., 2024)

• The fusion step constructs a temporal multi-graph $\tilde{A}_t^g$3 by unioning the edges of consecutive snapshots.
• Edge weights are determined by Hawkes process decay, e.g.,

$\tilde{A}_t^g$4

• The Hawkes-weighted adjacency is normalized and input to a GCN or GAT.

Multimodal Dynamic Fusion (MM-DFN) (Hu et al., 2022)

• Nodes are uniquely associated to modality-utterance pairs. Edge weights reflect cosine similarity.
• DMF applies, per fusion layer, an LSTM-style gating mechanism (update, forget, output gates) and a graph convolution over the dynamic graph-constructed features, with residual mixing:

$\tilde{A}_t^g$5

$\tilde{A}_t^g$6

The fusion of gate outputs with graph convolutions at each layer adaptively manages redundancy and complementarity.

4. Computational Properties and Training Procedures

DMF modules impose additional computational cost over single-graph methods, but are algorithmically tractable via:

• Light-weight attention or gating (softmax or LSTM-cell style gates) for fusion.
• Parallel per-graph GCN or GAT layers, followed by fusion.
• Efficient mini-batch and multi-stage training protocols (notably the 3-step mini-batch in SFDyG (Qi et al., 2024)) that decouple memory requirements from the number of input snapshots.

DMF frameworks generally optimize standard supervised or self-supervised losses (cross-entropy or MSE) with optional L₂ regularization on fusion-specific parameters.

5. Empirical Results and Ablation Findings

Ablation studies in multiple domains establish the efficacy of DMF modules over single-graph and static fusion baselines:

Experimental Setting	RMSE (lower is better)	Relative Improvement
Single-graph (distance/traffic)	457.4 / 453.3	Reference
DMF (fusion w/o RL)	448.0	~5–9 better
Full RL-DMF (fusion + RL)	426.4	~27 better

In SFDyG, DMF enables full-batch and scalable mini-batch link-prediction on large temporal graphs, with up to 50% GPU memory savings compared to per-snapshot or fully recurrent approaches (Qi et al., 2024).

MM-DFN demonstrates that dynamic gating and residual mixing in DMF yield enhanced complementarity and reduced redundancy compared to static stacking of GCN layers in multimodal emotion recognition (Hu et al., 2022).

6. Contrast with Static and Single-Graph Fusion

DMF modules systematically address three weaknesses of static or independent single-graph fusion strategies:

1. Redundancy Accumulation: Static GCN stacking may amplify irrelevant or redundant features across layers; DMF employs gates or attention to regulate information flow and suppress redundancy (Hu et al., 2022).
2. Lack of Adaptivity: Single-graph approaches cannot exploit heterogeneity between graphs or adapt to context; DMF enables data-dependent, often node-level, adaptivity (Rafi et al., 10 Jan 2026).
3. Computational Scalability: By jointly fusing multiple graphs or snapshots, DMF enables efficient training regimes that scale to long temporal windows or large multimodal interaction graphs, while retaining predictive accuracy (Qi et al., 2024).

7. Application Domains and Representative Implementations

DMF is deployed in diverse application settings, often as a critical architectural innovation:

• Evacuation Traffic Prediction: RL-DMF models fuse real-time distance and travel-time graphs, allowing robust, interpretable multi-horizon flow forecasting (Rafi et al., 10 Jan 2026).
• Dynamic Link Prediction: SFDyG fuses discrete-time graph snapshots with Hawkes process weighting, enabling scalable dynamic GNNs for future link prediction (Qi et al., 2024).
• Multimodal Emotion Recognition: MM-DFN leverages a DMF module for context-aware multimodal feature integration, controlling information redundancy and boosting emotion recognition under conversational settings (Hu et al., 2022).

These implementations establish the DMF paradigm as a flexible, domain-agnostic solution to fusing heterogeneous graph information dynamically and adaptively within end-to-end trainable architectures.