Date of Graduation
8-2025
Document Type
Thesis
Degree Name
Master of Science in Computer Science (MS)
Degree Level
Graduate
Department
Electrical Engineering and Computer Science
Advisor/Mentor
Nakarmi, Ukash
Committee Member
Brajendra Nath Panda
Second Committee Member
Lu Zhang
Abstract
A recommendation system is a bridge between users and products, which is widely used in e-commerce such as Amazon and Netflix. This study investigates the use of Graph Neural Networks (GNNs), Light Graph Convolution Network(LightGCN) and Graph Sample and Aggregate (GraphSAGE), in the recommendation system on two categories of Amazon review datasets ( "All Beauty" and "Tools and Home Improvement"). The novelty of this work includes combining supervised and self-supervised learning through Weighted Approximate Rank Pairwise (WARP) and Information Noise-Contrastive Estimation (InfoNCE) losses, to optimize the embeddings of users and recommended items in the shape of a ranking list. The results show that GraphSAGE outperforms LightGCN, and more specifically, among GraphSAGE models, the one with the combination of WARP and InfoNCE losses performs very well. The reasons behind these improvements could be that GraphSAGE leverages the node features to capture both collaborative and content-based signals. Moreover, it samples a fixed number of neighbors, not all of them, which reduces the noise’s impact, specifically in sparse and large datasets. GraphSAGE uses the ReLU activation function, which makes it capable of learning non-linear behaviors, such as price sensitivity. Besides the model, contrastive learning (InfoNCE) offers robustness and consistency by creating two augmented views of a graph. Optimization in InfoNCE aims to make these two graphs’ embeddings similar. Overall, this work contributes to recommendation systems through feature engineering, novel loss design, and scalability solutions for large-scale data. It can be shown that contrastive loss will improve this system’s performance significantly.
Citation
Aghamohammadghasem, M. (2025). Contrastive Loss in Recommendation Systems. Graduate Theses and Dissertations Retrieved from https://scholarworks.uark.edu/etd/5845
