Educational recommendation systems are essential for personalising learning experiences, as they help identify knowledge gaps and facilitate targeted interventions in the online learning environment. The efficiency of these systems depends on the quality of input data, especially effective data analysis, which begins with data preprocessing. Unfortunately, many developers tend to overlook this critical phase, concentrating instead on the modelling process or using an available pre-cleaned public dataset. In online education, this oversight can lead to misleading predictions, especially when dealing with heterogeneous, incomplete and context-sensitive educational data. This study evaluates the impact of two different data preprocessing pipelines applied to a single predictive model utilising the large-scale EdNet-KT1 dataset. We compared two distinct preprocessing strategies: a standard pipeline characterised by minimal cleaning and encoding and a context-aware pipeline that incorporated temporal, behavioural and difficulty-based features. The findings indicate that the context-aware pipeline resulted in a slight decrease in overall accuracy, which dropped from 0.93 to 0.80. However, it notably improved the recall for the minority class, increasing it from 0.25 to 0.89. This suggests that the context-aware pipeline demonstrates enhanced sensitivity to learning progress, particularly in recognising the minority class more effectively. Overall, the study underscores the need for thoughtful, context-aware preprocessing to build reliable and effective educational recommendation systems.