Artificial intelligence-enabled transaction prediction.

Predicting transaction behaviour and volume on e-commerce websites is difficult because transactional statistical methods don't do well with complex, unstructured data and cannot account for external factors like trends and promotions. Also, existing AI transactional predictive models suffer from historical data biases, interpretability concerns, and problems with human-crafted features. This research therefore proposes an AI machine learning transaction prediction system model that is more robust and transparent for predicting transaction behaviour when shopping online. This project designed a feedforward neural network (FNN) for binary classification to predict whether a transaction is made in the Santander customer transactions dataset. The input layer had 202 neurons, consisting of the data features, two hidden layers of 10 neurons each with ReLU activation, and a dropout rate of 0.2 to prevent overfitting. The output layer has 2 neurons with a sigmoid activation function to output separate probabilities, thus balancing complexity and enhancing class differentiation. The research made use of Bayesian optimization and tuned key hyperparameters, learning rate (0.001), batch size (32), drop rate (0.3), and optimiser (Adam), to balance regularization and convergence. The Adam algorithm combined with binary cross-entropy as the loss function was applied to a 70% training, 20% validation, and 10% split to improve stability and generalization. Our proposed approach demonstrated a high accuracy of 0.93, outperforming traditional models like decision tree (0.84), random forest (0.87), logistic regression (0.85), and gradient boosting (0.90). The Time-Weighted F1 Score of 0.52 and a Temporal ROC-AUC of 0.75 from time-dependent metrics are particularly important for real-world applications, as they reflect the model's capability to adapt to the evolving nature of transaction. The model development process is thoroughly documented, with a clearly defined architecture and an accessible implementation environment, thereby enhancing interpretability and helping to mitigate biases.

Grantor dc:publisher.institution

Robert Gordon University

Year dc:date.issued

2025

Author dc:creator

• Atalor, Daniel Osaroboh

Advisor dc:contributor.advisor

• Y. Zhao and S. Kannan

• Artificial intelligence
• Machine learning
• Neural networks
• Customer transactions
• Trend prediction

Language dc:language

en

Identifier

oai:rgu-repository.worktribe.com:2934768
https://doi.org/10.48526/rgu-wt-2934768

OAI identifier oai:identifier

oai:rgu-repository.worktribe.com:2934768