Dynamic Attention Mechanism for Automated Visual Affect Recognition: Spatiotemporal Learning models, Retail Deployment Considerations, and Causal Limitations

The facial emotion recognition (FER) system can be used to real-time analyse shoppers' emotions. But previous CNN-based FER approaches have not modelled temporal information and have overlooked issues on distribution shift from lab-based to retailers' setups, privacy concerns regarding biometric data processing, and the unjustified causal relationship between emotions and consumption. Temporal features were extracted using four different LSTM networks evaluated on CK+ (981 sequences from 123 subjects) under a subject-independent partition. The Attention BiLSTM achieved an accuracy of 92.42% on the evaluated CK+ temporal-sequence setting, outperforming the Deep BiLSTM by 6.06 percentage points while using fewer parameters. The results indicate that the evaluated attention-based architecture provided a more effective inductive bias than increased recurrent depth under the experimental conditions considered. A qualitative distribution-shift analysis across nine dimensions identifies substantial differences between the benchmark datasets and retail CCTV conditions. The resulting retail-domain performance remains an empirical question, because no retail video dataset was evaluated in this study. A privacy-preserving deployment architecture is proposed, based on on-device processing, short-lived image buffers, data minimization, and transmission of aggregate emotion outputs rather than facial images. Legal compliance is treated as deployment- and jurisdiction-dependent rather than as established by the architecture alone. A causal-analysis framework is presented to identify potential confounding and proxy-corruption mechanisms in future emotion-purchase studies. Causal effects are not estimated, because transaction-level purchase outcomes are not available in the present study. These insights provide the technical and real-world limits for retail analytics from emotional expressions.