Sentiment analysis and preference prediction of advertising images based on eye tracking

Abstract

In the advertising industry, understanding consumers’ sentiment reactions and preferences towards advertising images is crucial. To gain a deeper understanding for the impact of advertising images on audience sentiments and preference prediction, a saliency prediction model and an advertising preference prediction model for text reinforcement learning are proposed based on eye movement data. The saliency prediction model enhanced the visual features of advertising images by analyzing the text saliency in advertisements. The advertising preference prediction model utilized eye movement data and graph convolutional neural networks to understand individual visual reactions and sentiment preferences towards advertising images. The performance test results indicated that the proposed significance prediction model performed excellently on the accuracy-recall curve and F-measure curve. The accuracy on the model prediction performance of the study using the method was 0.787, SROCC was 0.628, PLCC was 0.583, and EMD decreased to 0.064. The highest SROCC value between the model and the true value can reach 0.6169, the highest PLCC can reach 0.5722, and the lowest error is 0.1929. The average similarity of the advertising preference prediction model was about 0.55 when the liking score was 1. When the score was 7, the average similarity increased to about 0.65. It presents that the method can effectively predict different preferences and meet practical application needs. As a result, the study not only enriches the research framework of advertising sentiment analysis and preference prediction but also provides more practical tools and methods for advertisement design, personalized recommendation, and marketing strategy development. Sentiment analysis models can be further explored in future research and their practical applications can be developed to promote the digital and intelligent transformation of the advertising industry.