Recommender system for online shopping

• Main Author: Wan, Tianyi
• Other Authors: Zhang Jie
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/153259
• Institution: Nanyang Technological University

Recommender systems are changing from novelties utilized by some E-commerce sites to important commercial enterprise tools. All the large E-commerce sites have their own recommender systems to recommend products to customer. Current recommendation algorithms usually learn the ranking scores of items by using a single task like conversion rate base on the user data. However, there are new ways such as adding transformers to model users’ multiple types of behaviour sequences like click through rate (CTR) and conversion rate (CVR) simultaneously. Many of the new recommender systems have also been using multi-task learning (MTL). Studies also found that there is a seesaw phenomenon on multi-task learning where performance of all tasks is not matching which means performance of a task is improved by sacrificing other tasks. However, there is not a single recommender system that model both the multiple types of user behaviour sequence and seesaw phenomenon of multi-task learning. In this paper, I proposed a Multi-transformer Progressive Layered Extraction (MTPLE) model for online recommender system. It utilises the transformers for user’s multiple types of behaviour sequences simultaneously. In addition, it uses progressive layered extraction to optimize multiple objectives and reduce seesaw phenomenon of multi-task learning. Experiments on JD RecSys Dataset were carried out to demonstrate the effectiveness of MTPLE. MTPLE achieved an overall improvement in performance compared to other models.