Answers are grounded in this paper and cite their sources.
Multimodal time-series classification has been successfully applied to diverse domains, yet recent research has increasingly favored complex feature-level fusion architectures whose computational cost often limits practical deployment. In this work, we investigate whether such complexity is necessary through a comprehensive cross-domain empirical study of decision-level fusion. We introduce MATIC (Multimodal Attention-based Time-series Classifier), a lightweight framework in which channel-specific convolutional neural networks independently classify each signal stream and a self-attention module adaptively fuses the resulting class-probability distributions. This modular formulation decouples feature extraction from multimodal fusion, enabling efficient training, reduced model complexity, and straightforward scalability across heterogeneous modalities. We evaluate the proposed approach on six public benchmark datasets spanning emotion recognition, human activity recognition, and stock movement prediction. Across all benchmarks, MATIC achieves competitive performance while requiring substantially fewer trainable parameters (0.11–1.95 M) than many specialized multimodal architectures. Comparisons with conventional decision fusion strategies and architectural ablations show that adaptive attention outperforms static and MLP-based fusion in most settings, and robustness analyses characterize its sensitivity to noisy and missing inputs. These findings suggest that lightweight decision-level attention fusion constitutes a strong baseline for multimodal time-series classification and provides a simpler alternative to increasingly complex multimodal fusion architectures.
No takes yet. Share an insight, caveat, or question.
Nepal et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: