Learning to Route: Per-Sample Adaptive Routing for Multimodal Multitask Prediction

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Linguistics

Background:

• Data heterogeneity and varying task interactions pose challenges in multitask, multimodal prediction.
• Existing models often use fixed strategies, failing to adapt to sample-specific data characteristics.

Purpose of the Study:

• To develop a unified framework for adaptive routing in multitask, multimodal prediction.
• To dynamically select modality processing pathways and task-sharing strategies on a per-sample basis.

Main Methods:

• Introduced a routing-based architecture with multiple modality paths (raw and fused text/numeric features).
• Learned to route inputs to the most informative modality-task expert combination.
• Employed shared or independent heads for task-specific predictions, trained end-to-end.
• Evaluated on synthetic and real-world psychotherapy notes for depression and anxiety prediction.

Main Results:

• The proposed method consistently outperformed fixed multitask and single-task baselines.
• The learned routing policy offered interpretable insights into modality relevance and task structure.
• Demonstrated effective per-subject adaptive information processing.

Conclusions:

• The adaptive routing framework effectively handles data and task correlation heterogeneity.
• This approach advances personalized healthcare by tailoring information processing to individual needs.
• The model provides a flexible and interpretable solution for complex prediction tasks.