Date of Graduation

5-2025

Document Type

Thesis

Degree Name

Bachelor of Science in Data Science

Degree Level

Undergraduate

Department

Data Science

Advisor/Mentor

Schubert, Karl

Committee Member

Laskar, Mohammad Nasir Uddin

Second Committee Member

Davis, Emmale

Third Committee Member

Tran, Quoc

Abstract

The expansion of e-commerce has continued at a blinding pace since the COVID-19 pandemic, and retailers are constantly looking for new ways to retain customers. Ensuring that diverse and well-classified images are on product pages has been a paramount method for retailers to ensure retention as they increase product engagement and sales and enhance user experience. Managing and labeling these vast catalogs of images by hand is becoming increasingly infeasible, so some online retailers have started to turn to automated classification models to assist them. Accuracy in these classification models is integral, as a good image classification model can improve search functionality and give retailers more accurate data on what types of products are on their site. In this study, we investigate the performance of various deep convolutional neural networks, specifically ResNet-50, EfficientNet-B0, and EfficientNet-B2, when it comes to classifying images in a hierarchical system designed for Walmart’s product catalog. To reduce training time and resource consumption, we utilized versions of these models that had been pre-trained on ImageNet and split classification models into parent and child modules. Our best-performing model was EfficientNet B2, which had an adaptive learning rate and achieved 91.83% top-1 accuracy on the parent labels and 88.57% accuracy on a 12-label child model. Performance was generally consistent with related literature, but some categories, particularly those dependent on specific product angles or the presence of text, underperformed in precision, recall, and F1 scores. This suggests that refining business definitions for image classification by reducing vagueness could improve model effectiveness. While our final results fell short of our initial 95% accuracy goal for both models, they showed significant potential for future scalable deployment. With further refinement, these models may support automated image classification pipelines and assist decision-making in forming e-commerce guidelines for product image consistency and labeling for Walmart.

Keywords

machine learning; image classification; artificial intelligence; ResNet; EfficientNet

Download

Included in

Data Science Commons

Share

COinS