The integration of contextual information, like time, weather, or location, into machine learning (ML) models has been shown to improve the performance and personalization of the model. However, these additional features may unintentionally introduce biases, leading to disparities across target classes or subgroups. This paper presents a novel methodology to evaluate the bias introduced by contextual features in ML models. We introduce a general metric, called Contextual Bias, which quantifies the disparity in class-level performance when each contextual feature is removed. To efficiently assess feature influence without retraining the model from scratch, we leverage DaRE (Data Removal-Enabled) forests, a machine unlearning framework that allows post-hoc removal of contextual features. This approach is applied to a real-world dataset of tourist visits to Points of Interest (PoIs) in Verona, Italy, where contextual factors play a crucial role in shaping user behavior. This work provides a foundation for developing unbiased context-aware systems, highlighting the need to consider not only accuracy but also bias metrics when integrating contextual information into ML models.