AI-Based Personalization Of Customer Experience Using Transformer Networks

Abstract

The sheer number of digital commerce platforms and the exponential increase in the volume of consumer-generated data have spurred the need for intelligent systems of personalisation that are contextually aware. The traditional collaborative filtering and content-based recommendation methods are increasingly insufficient to reflect the temporal variability, semantic complexity and affective aspects of today's interactions between customers. This paper presents an AI-based customer experience personalization framework called TransNet which combines the Transformer encoder network with sentiment analysis using Bidirectional Encoder Representations from Transformers (BERT). The architecture is designed to capture sequential user behavior patterns using multi-head self-attention mechanisms, temporal interaction dynamics through positional encoding, and emotional engagement through a dedicated sentiment analysis module which analyzes customer review text. TransNet combines behavioral signals with affective features to create enriched user representation vectors which are used as input to a downstream personalization engine, which is responsible for real-time product ranking and adaptive recommendation. The experiments are carried out on the Amazon Reviews 2023 dataset which is a collection of more than 571 million reviews from Amazon on 33 product categories collected from May 1996 to September 2023. The proposed model outperforms the BERT baseline with an accuracy of 95.3%, precision of 94.7%, recall of 93.8%, and F1-score of 94.2%. Ablation experiments show that each of the architectural components (positional encoding, multi-head attention, sentiment module) make a meaningful contribution to performance. The results show that Transformer-based models with sentiment-enhanced feature fusion are highly effective and scalable for AI-driven personalization in dynamic e-commerce settings, effectively managing customers' experience. The results indicate that the sentiment-augmented feature fusion with the Transformer-based architectures is a highly effective and scalable approach to customer experience personalization in dynamic e-commerce.