User feedback could give developer an information on what should be fixed or should be improved. But there are many user feedback that are actually spam. In user feedback, spam contents are more likely to be an inappropriate feedback, a feedback that is not actually a feedback, just some random comment or even a question. Reading and choosing feedback manually could be costly, especially in terms of time and energy. Therefore, this research focuses in building a spam filtering model using Multinomial Naïve Bayes that implement a TF/IDF approach to detect spam automatically. For text classification, Multinomial Naïve Bayes proved on having better speed and having good performance. With TF/IDF, word that highly occurred in many documents has less impact than other so it could help increasing performance from imbalanced dataset. This research aims to implement Multinomial Naïve Bayes for spam filtering in user feedback and to measure performance of the model. Best performance of this classifier was obtained when using up-sampling method and typo corrector with 70:30 ratio of train and test set resulting in 89.25% for accuracy, 45% for precision, 56% for recall, and 50% for F1-Score.