Multimodal user behavior analysis supports closed loop optimization of smart home interaction flows

Although smart home ecosystems have grown rapidly in scale and interaction complexity, bottlenecks in device pairing, scene automation, cross-device coordination, and feature discovery persist across major platforms. Existing usability research often relies on expert evaluation, questionnaires, or task-level summaries that lack the behavioral granularity needed to prioritize specific interaction nodes. Positioned as a Human–Computer Interaction and usability-engineering study, this paper proposes MAOV (Monitor–Analyze–Optimize–Validate), a conceptual closed-loop method for data-informed interaction-flow redesign. It integrates five synchronized behavioral data streams into a three-level User–Task–Node structure. An Interaction Bottleneck Index (IBI) enables node-level friction diagnosis, while an Adaptive Multi-modal Behavioral Fusion Network (AMBF-Net) with Bottleneck-Modulated Cross-Modal Excitation generates enhanced embeddings for clustering. K-means++ with Bootstrap Adjusted Rand Index validation segments users into behavioral archetypes to inform human-led redesign. A within-subjects experiment ( $$N = 40$$ ) evaluated Baseline and Optimized laboratory prototypes across seven tasks spanning four pain-point dimensions. The optimized prototypes produced consistent improvements: task completion time decreased by 30.8 to − 49.8%, error rate by 56.6–−64.8%, SUS improved from 61.45 to 78.10, and NASA-TLX decreased by 40.6%. AMBF-Net embeddings also strengthened the stability and interpretability of behavioral clustering. These results demonstrate the effectiveness of one complete MAOV cycle under controlled conditions and provide a basis for longitudinal and platform-specific field validation. The evidence supports framework effectiveness only in this controlled experimental setting; effectiveness and transferability on live commercial platforms and in occupied homes remain unverified and require field validation.