Tavsiye sistemlerinde iyileştirmeler: Veri seyrekliği ve sınıf dengesizliğinin öznitelik entegrasyonu ve ağ bilimi ile ele alınması

Recommendation systems suffer greatly from data sparsity and class imbalance, which reduces the quality of user recommendations and limits the system's capacity to use the entire item inventory. Addressing these issues is crucial for improving recommendation accuracy and delivering well-suited suggestions to users. This thesis introduces novel ways to improve recommendation accuracy by incorporating item attributes into user-item interaction and using network science in negative sampling in collaborative filtering (CF) algorithms. Firstly, we proposed a data granularity framework that systematically integrates item attributes at different levels of detail within the bipartite network structure. This enriches the item nodes with detailed information, hence enabling the CF algorithms to potentially minimize the negative effects of data sparsity on the recommendations. Secondly, we designed network-oriented negative sampling strategies that leverage user-item interaction structure while selecting more informative negative samples for the CF algorithm's learning task. Unlike traditional negative sampling, these strategies personalize the selection process for each user and prioritize negative items according to different network-related factors like node distance and link density. For the two problems mentioned, i.e. data sparsity and class imbalance, we tested our proposed solutions, i.e. the data granularity framework and network-oriented negative sampling, respectively, with large sets of experiments. The results show that our approaches mitigate the negative effects of the data sparsity and class imbalance, leading to more robust recommendations. These findings may pave the way for new developments in this field.