Abstract

Numerous AI-based solutions have been developed to predict online shopping behavior and cart abandonment, ranging from statistical and sequential data mining approaches to deep learning models. However, this task remains challenging due to the inherent class imbalance of online shopping data and the lack of transparency of most successful classifiers. In this paper, we investigate both issues in the context of cart abandonment prediction using imbalanced clickstream data in online shopping. As a first contribution, we study the effectiveness of addressing class imbalance via thresholding compared to other undersampling and oversampling in the context of clickstream data. Therefore the thresholding method is investigated that balances the decisions generated by deep neural networks by adjusting the probability threshold through binary search. The advantage of this strategy is that it neither discards any relevant data samples nor introduces synthetic ones. As a second contribution, we study the contribution of sequential and static features to the model's performance and investigate the extent to which oversampling techniques induce noise in the interpretability results. The numerical simulations show that our balancing strategy is slightly superior to undersampling and oversampling approaches without inducing a bias towards any of the decision classes. Moreover, we noticed that integrating static features alongside sequential data further boosts the performance of the recurrent neural network models, aligning with the results of previous research.