PREDICTIONS OF CUSTOMER BEHAVIOUR OVER ECOMMERCE WEBSITES AND ANTICIPATING THEIR INTENTION

Abstract

This paper presents a real-time behavioural analytics solution for online consumers that consists of two modules that estimate visitor purchase intent and website desertion probability at the same time. The first module predicts a visitor's purchase intent using aggregated page view statistics acquired during a visit as well as certain session and user information. The collected features are fed into classifiers such as random forest (RF), support vector machine (SVM), and multilayer perceptron (MLP). To increase classifier performance and scalability, use oversampling and feature selection pre-processing processes. The findings reveal that MLP calculated with a robust backpropagation method with weight backtracking outperforms RF and SVM in terms of accuracy and F1 score. Another discovery is that, although clickstream data gathered from online navigation patterns transmit vital information about visitor buy intent, session information bases include unique information about purchasing inclinations. In the second module, everyone estimate the likelihood of a visitor's desire to exit the website without completing a purchase using just sequential clickstream data. Everyone trains a long short-term memory-based recurrent neural network to provide a sigmoidal output indicating the predicted horizon. When used together, the modules detect visitors who are ready to make a purchase but are likely to depart the site within the forecast time and take the necessary steps to enhance website abandonment and buy conversion rates to do. Our findings indicate the viability of employing clickstream and session information data to forecast purchase intent in virtual retail environments in an accurate and scalable manner.