ONLINE SHOPPERS INTENTION PREDICTION USING MACHINE LEARNING ALGORITHMS

Abstract

In today's fast-moving e-commerce landscape, being able to accurately predict whether an online shopper is likely to make a purchase is incredibly valuable for businesses aiming to boost sales and stay ahead of the competition. While earlier machine learning models like XGBoost and Random Forest have shown decent results in this area, they struggle to capture the complex relationships between features or interpret sequential user behavior. Transformer-based models originally designed for natural language tasks have started gaining traction in structured data prediction due to their ability to model interactions and dependencies more effectively. This research takes an in-depth look at two such models: SAINT and FT-Transformer. When tested on a familiar dataset for shopper behavior, SAINT achieved an accuracy of 89.91% and an AUC-ROC of 90.65%, while FT-Transformer also gave the same accuracy but slightly lower AUC-ROC at 89.68%. When compared with traditional models like XGBoost and Random Forest, which are the same in accuracy, they fell short in AUC-ROC, highlighting the superior ability of transformers to deal with imbalanced datasets. The attention mechanisms in SAINT and FT-Transformer helped identify detailed patterns in user sessions, resulting in better generalization. These findings offer promising directions for more intelligent, data-driven marketing strategies.