Many banks categorize customer transactions, increasingly relying on machine learning techniques. User feedback is often treated as ground truth, however, people often seem to intentionally or unintentionally mislabel transactions, which impairs model performance. This study is the first to examine whether incorporating the reliability of user feedback improves performance in the categorization of payment requests. For this purpose, four reliability metrics were measured: entropy, inter-annotator agreement, user history, and semantic category distance. These were combined into a composite reliability score using principal component analysis (PCA), after which four XGBoost models were trained: a baseline model that does not account for reliability, a model that excludes unreliable feedback from training, a model that weights feedback based on reliability, and a model that combines both filtering and weighting. Model performance will be evaluated using the weighted F1-score and the average inference time per transaction in milliseconds.