Cognitive biases in digital food advertising and their emotional impact: a comparative approach based on artificial intelligence and computer vision models

Abstract

Digital food advertising systematically exploits perceptual visual biases to influence both emotions and consumer behavior; however, the existing body of research lacks computational frameworks that link the automated detection of visual biases with probabilistic emotional inference. The purpose of this article is to design and validate a hybrid architecture for inferring probable emotional profiles based on visual biases in food advertising images, that is, a combination of Deep Computer Vision techniques and probabilistic modeling. To this end, a two-module architecture was constructed: the first is based on a Deep Computer Vision model with multi-label classification, using the EfficientNet-B0 model, which was compared with the ResNet-50, ViT-B/16, and MobileNet-V3 Large models; the second was the corresponding probabilistic component based on a Markov chain that maps the detected bias samples into ten emotional categories relevant to food marketing. A dataset was also created comprising 1,375 food advertising images from social media, which were manually annotated across eleven established categories of visual bias. Regarding the results, the EfficientNet-B0 architecture achieved the following scores: mean Average Precision = 0.961, ROC-AUC = 0.968, and weighted F1 score = 0.898, outperforming the architectures implemented in this study, while the Markov-inspired affective module produced emotional profiles that are easily interpreted in empirical literature. Finally, the conclusion is that the proposed architecture decouples the perceptual level from the affective level, enabling mechanistic explainability, updates via modules, and the extensibility of the emotional space, and allowing for practical applicability in marketing research, public health, and the regulation of commercial communication.