Research on Personalized Marketing of E-commerce Based on Transformer Architecture

Objective: Aiming at the problem that the traditional recommendation system is difficult to capture the user's dynamic interest and long-term behavior dependence, this paper explores the optimization method of e-commerce personalized marketing based on Transformer architecture. Methods: a multi-task Transformer model was constructed to globally model user behavior sequences using self-attention mechanism, and the Amazon dataset was extended with data augmentation techniques to jointly optimize the indicators of click-through rate (ACTR), conversion rate (ACVR) and coverage rate (ACOV). Results: In the test set, the Transformer model was significantly ahead, with ACTR (0.63), ACVR (0.47), and ACOV (0.78) being 23.5%, 38.2%, and 25.8% higher than optimal baseline, respectively, verifying its ability to balance and optimize the dynamic evolution of user interest and long-tail exposure. Conclusion: Transformer breaks through the timing constraints of traditional models through global dependency modeling, and provides a solution for e-commerce recommendation with both accuracy and ecological health. The dynamic weight allocation mechanism can simultaneously improve user attractiveness, transaction conversion and category coverage.