Consumer Purchase Intention Prediction Model Based on Multimodal Emotional Data Mining

Abstract

This paper proposes a dual-modal fusion model with cross-modal attention for consumer purchase intention prediction. The model integrates textual semantic features extracted by DistilBERT with structured behavioral features derived from review text, enabling cross-modal feature interaction through a Cross-Attention mechanism. Experiments conducted on 75,000 samples merged from the Yelp Review Full and IMDB datasets show that the model achieves an accuracy of 0.9852, a Macro-F1 of 0.9850, and an AUC-ROC of 0.9991, improving Macro-F1 by 2.10 percentage points over the text-only baseline and by 10.50 percentage points over the SVM baseline. Results validate the effectiveness of the multimodal fusion strategy for purchase intention prediction.