<![CDATA[Endless Runner games feature exponentially increasing difficulty as distance grows, often causing character failure and player frustration, largely because players struggle to select power-ups suited to their current difficulty context. While prior recommender-system research has mostly focused on purchase prediction for monetization, this study instead builds a personalized item recommendation system aimed at reducing failure and maximizing scores. We propose a hybrid Neural Collaborative Filtering (NCF) architecture combining General Matrix Factorization (GMF), which captures linear preferences, with a Multi-Layer Perceptron (MLP), which models non-linear interactions between playstyle and failure context (cause of death). The model was trained on 10,000 gameplay activity logs containing features such as jump count, obstacles avoided, and death cause. Over 20 training epochs, both training and validation accuracy converged to approximately 0.88–0.90, with a negligible gap between the two curves, indicating minimal overfitting. These results demonstrate that integrating GMF and MLP effectively produces recommendations adaptive to dynamic gameplay conditions.]]>
Copyrights © 2026
