İşbirlikçi filtreleme performansının çok boyutlu analizi: Metrikler, topolojiler ve algoritmalar

Recommendation systems play a crucial role in digital platforms by offering relevant items to users based on their interaction histories. Although considerable progress has been made in developing new algorithms, evaluating these systems remains challenging due to the overreliance on accuracy metrics and limited understanding of how dataset characteristics influence model performance. This thesis proposes a comprehensive evaluation framework for personalized collaborative filtering models (PCFMs) through a systematic analysis of 13 recommendation algorithms across 10 real-world bipartite datasets, using 12 performance metrics spanning accuracy, diversity, and fairness. Our experimental design is structured around four research questions, examining performance variations across evaluation dimensions, the coherence among metrics, model sensitivity to metric types, and the impact of graph topological features on performance. The results show that graph-based models achieve more balanced outcomes across all three evaluation aspects. Additionally, while accuracy metrics exhibit strong internal correlations, diversity and fairness metrics display more complex and model-dependent relationships. Moreover, topological features such as sparsity and L3 similarity significantly affect performance, though their influence varies by metric and model type. Overall, this study demonstrates that reliable evaluation of recommendation systems requires a multidimensional perspective, offering insights that are essential for both academic benchmarking and real-world deployment.