Graph-based recommendation algorithms are increasingly popular, but concerns remain about biases and the impact of user behavior on performance. This study addresses two key questions: Are recommendations independent of item interaction frequency, and how does user activity influence model effectiveness? We evaluate multiple models across diverse datasets using fairness, diversity, and accuracy metrics. Results show that while recommendations are not entirely independent of item degree, advanced techniques like data augmentation and contrastive learning significantly reduce popularity bias compared to traditional methods. User activity levels also substantially affect performance, with impacts varying by dataset structure. These findings highlight the complex interplay between user behavior, dataset characteristics, and model performance, emphasizing the need to balance accuracy and fairness in recommendation systems. Model selection should be guided by dataset structural properties.