Source-linked AI summary
MURAL: Multimodal Uncertainty-aware Recommendation via Adaptive edge Learning
Ahmad Mousavi, Majid Alikhani, Yeon-Chang Lee, Roberto Corizzo, Yeganeh Abdollahinejad
TL;DR
Multimodal recommendation is limited by static graph topologies and indiscriminate fusion of noisy modality signals. MURAL learns adaptive item correlations, models modality uncertainty, and uses gradient-detached behavioral anchoring; across large-scale benchmarks, it consistently outperforms structural and generative state-of-the-art models.
Problem
Existing multimodal graph recommenders rely on static topologies and deterministic fusion, limiting adaptation to evolving semantics and heterogeneous modality reliability.
Method
MURAL combines retrieval-augmented Adaptive Edge Learning, aleatoric uncertainty-aware fusion, and gradient-detached alignment to behavioral signals.
Results
MURAL consistently outperforms structural, generative, and alignment-based state-of-the-art baselines across multiple large-scale benchmarks.
Takeaways & Limitations
Dynamic topology learning and reliability-aware fusion provide an interpretable and efficient framework for robust multimodal recommendation.
Takeaways & Limitations
Future work should model user-specific modality preference variance and examine how adaptive semantic neighbors shift over longer training horizons.
Abstract
from arXiv · showhide
Multimodal Graph Neural Networks have become standard for recommendation by augmenting sparse interaction data with content features. Yet current architectures face two bottlenecks: structural rigidity, from a reliance on static precomputed similarity graphs that cannot adapt to evolving preferences; and semantic fragility, where noisy modality signals are indiscriminately fused, distorting the collaborative signal. We propose MURAL (Multimodal Uncertainty-aware Recommendation via Adaptive edge Learning), a unified framework that shifts multimodal recommendation from fixed structural augmentation to dynamic topology discovery. To address structural rigidity, an Adaptive Edge Learner combines a differentiable retrieval-augmented strategy with an approximate nearest neighbor search to discover latent item-item correlations that are both semantically adaptive and computationally scalable (O(NlogN)). To address semantic fragility, an Uncertainty-Aware Fusion module models the aleatoric uncertainty of heterogeneous modalities, dynamically down-weighting unreliable features while prioritizing high-confidence signals as a defense against cross-modal noise. We further employ a contrastive teacher-student alignment that anchors modality-specific representations to stable behavioral signals, ensuring optimization stability without gradient leakage. Experiments on large-scale benchmarks including TikTok and Amazon show that MURAL significantly surpasses both structural and generative state-of-the-art baselines, achieving superior accuracy while offering interpretability through domain-specific modality dominance and robustness under extreme data corruption.
1 Introduction
Multimodal recommendation faces structural rigidity from static graph construction and semantic fragility from indiscriminate fusion of unreliable modalities. MURAL addresses these bottlenecks through adaptive topology discovery, uncertainty-aware fusion, and gradient-detached behavioral-semantic anchoring.
- MGNNs can mirror the interaction graph, missing latent item-item correlations that exist within multimodal space.
- Static similarity heuristics impose structural rigidity by assuming the optimal semantic topology is fixed beforehand.
- Deterministic fusion creates semantic fragility by weighting noisy and high-fidelity modality features alike, potentially distorting collaborative signals.
- MURAL’s Adaptive Edge Learner discovers evolving semantic correlations differentiably while retaining O(N log N) scalability through retrieval-augmented search.
- Its uncertainty-aware fusion down-weights noisy or sparse signals and anchors representations to discriminative modality features.
- MURAL adds gradient-detached behavioral-semantic anchoring and consistently outperforms structural, generative, and alignment-based state-of-the-art baselines across large-scale benchmarks.
2 Related Work
Multimodal recommendation research progressed from hybrid feature integration to graph propagation and self-supervised alignment, but existing methods remain constrained by fixed topologies and uniform modality reliability. MURAL responds with dynamic topology discovery and uncertainty-aware alignment.
- Early multimodal recommenders combined visual, textual, and structural features before GNNs enabled graph-based multimodal propagation.
- Interaction-centric methods remain limited by the observed user-item graph, especially when interactions are sparse.
- GSL methods such as LATTICE and MMGSL mine item topologies from precomputed similarities but inherit heuristic neighborhood bias and uniform modality reliability assumptions.
- Multimodal SSL frameworks align features with collaborative signals but remain dependent on static item topologies and indiscriminate alignment pressure.
- MURAL replaces fixed heuristics with dynamic topology discovery and filters contrastive signals through uncertainty-aware fusion.
3 Adaptive Edge Learning with Uncertainty-Aware Fusion
MURAL learns adaptive item-item topology and reliability-aware multimodal representations instead of relying on fixed similarities and deterministic fusion. Its retrieval-and-refine graph construction combines scalable candidate search with learned edge weighting, while uncertainty modeling and behavior alignment address modality noise.
- Framework overview: MURAL combines adaptive topology discovery, modality dominance learning, aleatoric uncertainty modeling, and behavior-aligned representations in one differentiable framework.The framework transitions multimodal recommendation from static structural augmentation to dynamic, reliability-aware discovery.
- Behavior alignment: Behavior-aligned modality representations blend content embeddings with gradient-detached behavioral embeddings, preventing alignment gradients from distorting the interaction encoder.The coefficients controlling behavioral and content information are constrained to avoid shortcut learning, with the behavioral coefficient updated through the structural contrastive loss.
- Retrieval-augmented neighborhoods: O(N_i log N_i) complexity makes periodic approximate-nearest-neighbor re-indexing practical for large-scale recommendation ecosystems.Candidate sets remain fixed between refreshes, while edge strengths are dynamically reweighted; re-indexing every 10 epochs is reported as a trade-off between staleness and overhead.
- Adaptive edge learning: The Adaptive Edge Learner retrieves candidate neighbors, scores them with an MLP using multimodal and behavioral features, and retains the top-k_m neighbors in sparse modality-specific graphs.Local softmax normalization produces edge weights, while entropy regularization discourages one-hot neighbor distributions.
- Uncertainty-aware fusion: Uncertainty-aware fusion down-weights unreliable per-item, per-modality signals by modeling positive aleatoric variance alongside modality importance.Learned log-variance parameters and a regularizer keep uncertainty numerically stable and prevent variances from collapsing to zero.
- Graph propagation: The final propagation mixes normalized user-item and modality-specific item-item adjacency matrices with learnable nonnegative coefficients before LightGCN-style updates.After L propagation layers, each node representation is obtained by averaging the representations from all layers.
4 Evaluation
MURAL is evaluated on three multimodal recommendation benchmarks using full ranking and comparisons with established baselines. Results show consistent gains, while ablations and sensitivity analyses attribute performance to adaptive topology learning, uncertainty-aware fusion, contrastive alignment, and gradient detachment.
- 4.1 Experimental Setup: MURAL is evaluated on TikTok, Amazon-Baby, and Amazon-Sports using Recall@K and NDCG@K under full ranking.Baseline implementations were re-created in a unified environment with hyperparameter sweeps, and results were averaged over five runs.
- 4.2 Performance Comparison: MURAL consistently outperforms existing models, including recent multimodal graph structure-learning approaches, on Recall@20 and NDCG@20.The reported comparison attributes these gains to retrieval-augmented topology learning and uncertainty-aware fusion under multimodal noise.
- 4.2 Performance Comparison: MURAL surpasses MMSSL, DiffMM, and DiffCL by enriching the graph with latent item-item edges instead of relying on computationally intensive stochastic denoising.The paper presents this structural approach as a more interpretable and efficient pathway for mitigating data sparsity.
- 4.2 Performance Comparison: MURAL’s learned topology and aleatoric uncertainty modeling distinguish it from static graph construction and deterministic noise handling in MMGSL and FREEDOM.The topology is discovered during training, while uncertainty modeling suppresses item-level noise across heterogeneous datasets.
- 4.3 Ablation Studies: UAF prioritizes visual signals while attenuating textual and acoustic channels when they are sparse or non-discriminative.The TikTok analysis links this behavior to visual information density, background acoustic signals, and sparse textual descriptions.
- 4.3 Ablation Studies: UAF increases uncertainty for corrupted visual features, while clean visual features converge toward lower uncertainty and higher confidence.The fusion analysis further reports that attention and uncertainty scores provide distinct, uncoupled mechanisms for weighting modalities.
- 4.3 Ablation Studies: Removing gradient detachment reduces TikTok Recall@20 from 0.1221 to 0.1091, supporting behavioral-gradient isolation during modality alignment.The reported degradation occurs across all datasets and is interpreted as evidence that gradient leakage distorts the collaborative topology.
- 4.4 Hyperparameter Effects: Increasing neighborhood sparsity k initially improves Recall@20, but performance declines beyond an optimal threshold as weak item correlations introduce semantic noise.Performance remains stable near the peak, which the analysis associates with uncertainty-aware weighting of modality signals.
5 Conclusion and Future Work
MURAL addresses structural rigidity and semantic fragility through adaptive topology discovery and uncertainty-aware multimodal fusion. The authors report improved robustness and accuracy, while identifying user-side personalization and temporal graph dynamics as future directions.
- MURAL replaces static similarity heuristics with adaptive topology discovery through the Adaptive Edge Learner.The framework is presented as a response to non-adaptive graph construction and cross-modal noise.
- The Uncertainty-Aware Fusion module models aleatoric uncertainty to prioritize high-fidelity modality signals over unreliable features.This moves the framework beyond deterministic fusion.
- MURAL consistently outperforms state-of-the-art structural and generative models in empirical evaluations.The reported outcome is improved recommendation accuracy and robustness across the evaluated settings.
- Future work includes extending uncertainty-aware modeling to user-specific modality variance.This could personalize the fusion process beyond the current mechanism.
- Future work also includes studying how adaptive semantic neighborhoods shift over longer training horizons.This targets the temporal dynamics of the learned graph and evolving item trends.