CNN-based Predictive Analytics Models for ERP-Supported E-commerce

The rise of online shopping has pushed businesses to adopt advanced technologies and creative solutions to manage data effectively. ERP systems have become a cornerstone for many organizations, providing seamless integration, reducing repetitive data entry, and improving resource efficiency. Workflows, capacities, and company performance are all improved by these platforms. This paper introduces a machine learning (ML) and deep learning (DL) predictive analytics framework for e-commerce ERP systems. Our experiments used the Fashion-MNIST dataset, which contains 10,000 testing samples and 60,000 training samples. The dataset was thoroughly pre-processed, which included using Convolutional Neural Networks (CNN) to extract features, implementing logarithmic transformations, cleaning noisy data, and normalizing the data. Initially, we trained a sequential CNN model with multiple convolutional layers over 50 epochs, achieving a test prediction accuracy of approximately 91%. Our validation accuracy and loss analyses showed that the model was overfitting. To address this, we added Dropout layers to the model, which successfully reduced overfitting. After retraining with the same number of epochs, the improved model achieved an accuracy of around 93%. Despite this improvement, a few classes, such as Class 6 (Shirt) and Class 2 (Pullover), were not consistently classified correctly. This highlights some of the challenges in fine-tuning the model for specific cases. Overall, this study demonstrates the potential of CNN-based predictive analytics to enhance ERP-supported e-commerce systems and offers insights into overcoming challenges like overfitting and class misclassification. To further enhance the prediction process, we propose ensemble learning, class-balancing (e.g., SMOTE), and hyperparameter tuning (e.g., Bayesian Optimization). To further enhance the prediction process, we propose ensemble learning, class-balancing (e.g., SMOTE), and hyperparameter tuning (e.g., Bayesian Optimization). The novelty of this work lies in the integration of CNN into ERP-supported environments and the empirical validation using tailored CNN architectures that address SKU-level image classification needs and ERP system performance optimization.