Validated Multi-Modal Empathy Detection via Rank-Preserved Graph Fusion and Hybrid Attention Optimizations

HCI requires empathy detection; hence, multimodal learning frameworks that read language, audio, and visual inputs have been developed for this process.Time misalignment across modalities, unstable feature priority rankings, inefficient fusion algorithms, and real-world interpretability remain challenges for existing systems despite Temporal misalignment across modalities, instability in feature importance rankings, redundancy in fused representations, inefficient model optimization under computational constraints, and uncertainty-aware evaluation are addressed by an end-to-end multimodal empathy detection framework.The suggested analytical pipeline includes segment-level tri-stream contextual alignment, rankpreserving heterogeneous feature graphization, optimal-transport-based fusion, hybrid evolutionary-gradient attention optimization, and conformal prediction-driven calibration.A boundary-anchored contextual alignment system synchronizes text, audio, and video for segment-level conversational coherence.To stabilize importance ordering and convey crossmodal interdependence, aligned multimodal descriptors are structured into a rank-preserving heterogeneous feature graph.An optimal-transport-driven fusion method chooses and fuses features within computational limitations, removing redundancy and preserving semantically dominating cues.We combined representations with parameters and attention allocation controlled by a dual evolutionary gradient optimizer in our hybrid attention-driven convolutional architecture.Conformal prediction calibrates uncertainty-aware choice thresholded empathy evaluations.A large multimodal empathy corpus of healthcare consultations, consumer service encounters, and peer support chats evaluates the strategy.Segment-level experiments indicate steady alignment accuracy, ROC-AUC, and calibration reliability over three competitive baselines.The suggested method achieves 89.6% accuracy and a ROC-AUC of 0.935 on the entire test split while reducing inference cost to 19.9 GFLOPs within budget.These findings establish a technically coherent, interpretable, and resourceefficient pipeline for real-time empathy detection in diverse interaction scenarios.