Author: Roman Grebnev
Supervisor: Julia Neidhardt, Co-Supervisor: Ahmadou Wagne
Abstract
Online platforms often recommend items to users based on their current browsing activity, especially when long-term user history isn’t available. These Session-Based Recommender Systems typically rely on the sequence of items a user interacts with. However, just looking at the sequence (like a graph of interactions) or just looking at item descriptions (text) alone can be limiting. This master thesis explores how to improve these recommendations by effectively combining both types of information: the relationships between items (graph data) and their textual descriptions. We systematically investigated different strategies for merging these data sources – specifically, combining them early in the process (per item), later (per session), or using a simpler method of injecting text information into the graph analysis. Using experiments on two real-world e-Commerce datasets (AICrowd and Geizhals), we compared these combined approaches against systems using only graph or only text information. We assessed not only prediction accuracy but also computational efficiency and other qualities like recommendation diversity. Our findings show that combining the information after processing the graph and text data separately (session-level fusion) consistently gave the most accurate predictions. While combining earlier (item-level fusion) was computationally cheaper, the simple text injection method was less effective. We also found trade-offs between accuracy, computational cost, and the variety of recommended items. This research highlights the importance of how information is combined in Session-Based Recommender Systems and provides insights for balancing accuracy with practical constraints. We also deliver an open-source framework, SBRSFuse, to aid future research in this area.