Hedonic and utilitarian values in consumer interactions with LLM-based AI agents: a marketing perspective

An Adaptive Sparse Graph Neural Network (AS-GNN) framework is proposed for real-time classification of hedonic and utilitarian responses in neuromarketing applications, addressing the computational inefficiency of conventional multimodal fusion methods. The framework dynamically prunes non-critical cross-modal interactions using a gradient-weighted sparsification criterion, enabling latency-optimized processing of heterogeneous neurometric and behavioral data streams, including EEG, eye-tracking, and LLM-generated logs. A gated cross-modality attention mechanism selectively preserves high-impact edges for discriminative classification, while a modality-specific pruning controller adaptively regulates sparsification thresholds to maintain predictive accuracy under varying signal conditions. The proposed approach reduces computational overhead by approximately 63% relative to dense GNN architectures at 70% sparsity while maintaining classification accuracy above 92% (F1-score) at this operating point, on a synthetic dataset designed to emulate established neuromarketing signal properties, with real-world validation identified as a necessary next step, as validated through hardware-accelerated implementations with end-to-end latency below 50 ms. Furthermore, the integration of sparsity-aware adjacency reweighting and reinforcement-learned pruning policies ensures robust performance in dynamic shopping environments where consumer responses evolve rapidly. This framework bridges the gap between real-time processing constraints and the high-dimensional nature of neuromarketing data, providing a scalable solution for adaptive LLM-mediated consumer engagement. The modular design of AS-GNN enables seamless replacement of static fusion layers in existing pipelines, suggesting a practical integration path pending real-world validation.