A Transformer-Based Multi-Domain Recommender System for E-commerce

Abstract

Recommender systems are one of the most critical applications of AI, data science, and advanced analytics techniques because it has become integrated into our daily lives. Additionally, it serves as a powerful tool for making informed, effective, and efficient decisions and choices across a wide range of items. However, traditional techniques such as content-based and collaborative filtering often fail to consider the dynamic and short-term preferences of users when generating recommendations. To address this limitation, this research focuses on a session-based recommendation task using an XLNet transformer with various training strategies based on language modeling. Moreover, a dataset containing 102 million reviews of Amazon products was pre-processed to create two new datasets, one for a single domain and another for multi-domain data. A comparison between a GRU and the training strategies of XLNet reveals that the best training strategy achieves a 136.23% improvement in NDCG@20 and a 95.69% increase in Recall@20 for multi-domain data. In a single domain, it achieves a 168.81% improvement in NDCG@20 and a 25% increase in Recall@10.