Emotion recognition that combines EEG and eye movement could support mental-health monitoring and adaptive interfaces. But according to a new arXiv preprint, most fusion models treat the two signals as parallel features and classify each time window independently, so they degrade sharply when eye-tracking data is missing or unreliable.
The authors propose a Context-Aware Topological Distillation Graph Network. Instead of simply concatenating features, the model builds a graph that captures contextual and topological relationships between the signals. The "asymmetric" aspect means it is designed to keep working even when one modality—typically eye movement—is absent, rather than requiring both inputs at every step.
The abstract does not include experimental results, so the claimed robustness is based on the architecture's design rather than reported benchmarks. Still, the work points toward a more practical direction for multimodal emotion recognition, where sensor dropouts are common outside the lab. The preprint is available on arXiv under identifier 2610.26084.