Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Katerina Potika

Second Advisor

Sayma Akther

Third Advisor

Faranak Abri

Keywords

Recommender Systems, Graph Neural Networks, Hypergraphs, Knowledge Graphs, LightGCN, TransE, Link Prediction, Collaborative Filtering

Abstract

A recommendation system is an integral part of a wide range of businesses, such as E-Commerce, Social Networking Platforms, and Entertainment. However, existing graph-neural-based approaches still struggle with two main challenges: cold start (new items/users) and sparse interaction histories between users and items. Traditional Models, such as LightGCN, treat the user-item interaction history as a bipartite graph and learn embeddings by aggregating over pairwise edges rather than over real-world higher-order interactions. The question this project takes up is how to model higher-order relationships present in real user behavior so that item and user embeddings capture both these structural signals (hypergraph) and the semantic context of items (Knowledge Graph). The proposed approach, HyperKGRec, includes the following: each user is treated as a hyperedge connecting all the items they have interacted with, and a Knowledge Graph built from items’ metadata. Four progressive variants are evaluated on the Amazon Reviews 2023 Books dataset. The results show that usage of this higher-order structure, as well as Knowledge Graphs improves the performance of recommender systems in comparison to the baseline models.

Download

Available for download on Saturday, May 22, 2027

Share

COinS