E Commerce Consumer Behavior Prediction Model Integrating Spatiotemporal Features and Multimodal Sentiment Recognition

Abstract

The prediction of e-commerce consumer behavior is a complex task, traditionally relying on historical purchase data, which fails to fully capture the dynamic nature of consumer decisions. This study proposes a deep learning-based consumer behavior prediction model that integrates spatiotemporal features and multimodal sentiment recognition, addressing key limitations of traditional approaches. The model incorporates Long Short-Term Memory (LSTM), Transformer, and Graph Neural Networks (GNN) for spatiotemporal feature modeling, combined with BERT, ResNet, and Wav2Vec for sentiment analysis across text, images, and speech. To integrate these diverse modalities, an Attention-based fusion mechanism is employed to ensure optimal interaction and complementarity between the features. Experimental results show that the proposed model outperforms traditional baseline methods, such as Collaborative Filtering, LSTM, and BERT4Rec, in terms of accuracy, F1-score, and AUC-ROC. Notably, the proposed model improves AUC-ROC by 5.5%, confirming the importance of incorporating sentiment data and spatiotemporal information in consumer behavior prediction. The findings not only improve the accuracy and personalization of recommendation systems in e-commerce but also open doors for applications in fields like finance, social media, and beyond. This research contributes novel insights into the fusion of multimodal data and spatiotemporal behavior modeling, advancing the understanding of consumer decision-making processes.